Method, system, neural network, and computer program for executing computational tasks on multiple computing units

Abstract

The invention relates to a method, system, a neural network (DNN), in particular a pre-trained generative transformer (GPT), and a computer program for performing computational tasks on one or more computing units. The procedure (V, V1, V2) comprises: - Training (S20), in particular retraining (S20'), of a neural network (DNN, GPT) to generate a plurality of subtasks (R1, R2, R3) from a computational task (R) (S40), such that the one or more generated subtasks (R1, R2, R3, ...) are executable (S70) on multiple computational units (10, 11, 12, 13) to achieve (S70) a result (E, E1, E2, E3, ...) of the computational task (R). Alternatively or additionally, the procedure (V, V1, V2) comprises: - Generating (S40) a number of subtasks (R1, R2, R3, ...) from a computational task (R) depending on the output layer of a neural network (DNN, GPT); and distributing (S50) subtasks (R1, R2, R3, ...) from the number of subtasks (R1, R2, R3, ...) to several computing units (10, 11, 12) to achieve a result (E, E1, E2, E3, ...) of the calculation task (R).

Description

The invention relates to a method, system, a neural network, in particular a pre-trained generative transformer (GPT), and a computer program for executing computing tasks on a plurality of computing units.As various technologies progress, the computational tasks to be managed become more and more complex. These require ever larger (computing) resources. The corresponding hardware (in various embodiments) and / or energy can entail very high costs, an ecological burden and other disadvantages, which slow down the technological progress. Furthermore, the production or purchase of the hardware requires valuable (partly non-regenerative or limitedly available on the earth) substances (raw materials, valuable substances). This and a corresponding, overall high power consumption also impair the ecological balance of the technological progress in various sectors.The present document is concerned with the technical task of improving the benefit of computing units (of various types), in particular of hardware, in particular the cost-benefit ratio and / or ecological balance of computing units, in particular of the hardware or individual computing resources, and / or of enabling improved feasibility of computing tasks. In particular, the technical task includes enabling improved, as efficient and / or more reliable and / or more real-time processing of a computing task, in particular an external computing task (provided from the outside) and / or a computation task that is not known a priori.The object is achieved by each of the independent claims. Advantageous embodiments are described inter alia in the dependent claims. It is pointed out that additional features of a claim dependent on an independent claim can form a separate invention which is independent of the combination of all features of the independent claim without the features of the independent claim or only in combination with a subset of the features of the independent claim and which can be made the subject matter of an independent claim, a divisional application or a subsequent application. This applies in the same way to technical teachings described in the description, which may form an invention independent of the features of the independent claims.According to a first aspect, a (first) method for executing a computing task on a plurality of computing units (in particular also to be understood as being by means of a plurality of computing units) is provided. The (first) method comprises: forming a number of sub-tasks from a provided computing task as a function of the output layer of a neural network (in particular also to be understood as: as a function of the operation of the neural network); and (thereupon, as a function of the result of the forming of the sub-tasks): distributing the (one or more of the) formed sub-tasks from the number of sub-tasks to a plurality of computing units in order to achieve a result of the computing task.In this case, these processing units selected on, preferably depending on the output layer of the neural network, can be distributed (in particular also to be understood as being assigned in each case).Alternatively or additionally, the (first) method can comprise: distributing, from the number of sub-tasks formed in particular by the neural network, sub-tasks to a plurality of computing units from a number of computing units (e.g. configured and / or available computing units suitable for this purpose) for achieving a result of the computing task as a function of the output layer of a neural network. Alternatively or additionally, this can be applied to separate computing tasks. Particularly preferably, the forming of the sub-tasks takes place in such a way that they can be distributed or are distributed (e.g. for improved compliance with the target values described in the present document. This neural network can be one or more features described herein, in particular (preferably) the (same) neural network with which the calculation tasks are formed.The one or more of the plurality of computing units are in particular (in each case) computing units which (also or mainly) serve other purposes than executing the computing task. In particular, these are processing units which (in each case) serve different apparatuses, in particular different and / or in each case different purposes.Such purposes may be, for example, a general purpose of a universal data center, a specialized server, a specialized entity, etc. Alternatively or additionally, these may be devices or hardware that are used differently or specialized for different purposes, in particular optimized and / or (heavily, significantly) unevenly over time.For example, one or more of said plurality of computing units (each) belongs to: a server, PC (= Personal Computer), user equipment (e.g. smartphone, tablet, smart watch, smart close), electronic personal assistant, control equipment of a vehicle. At least two of the plurality of computing units are preferably of different types, in particular of the types specified above.One or more of said plurality of computing units or the device comprising the computing units are stationary devices. The plurality of the computing units can be located or operated in different time zones.In particular, one or more of the devices can be vehicles or one or more of the plurality of computing units can be those of the vehicles. Accordingly, these may (necessarily essentially) serve the purpose of executing the functionalities of the respective vehicle.The computing task is in particular a foreign device and / or a device read in from a further device and / or serves a different purpose than the devices comprising one or more computing units or the computing units.The computing tasks can relate to the handling of multimedia contents, in particular audio and / or video staring, coding methods, decoding methods, handling blockchains, in particular mining, generating or handling cryptodata and / or cryptocurrency, controlling robots and / or production steps.The computing units or the devices comprising the computing units can be located at different users and / or can be used by different users (optionally for different purposes). The term user is also to be understood as meaning a user or user: inside.The distribution can preferably (also) take place during the execution of the sub-tasks. In particular, the distribution also comprises (optionally dynamic) forming and / or redistribution of sub-tasks.The distributing may comprise or be a (e.g. confirmation-based) transmitting of the data of the sub-tasks and / or receiving of a confirmation, one or more intermediate results and / or results of the execution of the sub-task.The plurality of computing units (or the computing units available for selection) can be more than 2, 3, 5, 7, 9, 10, 50, 100, 1000 computing units. These may belong to the system described in the present document (in each case permanently or at least temporarily). This is also to be understood in particular in such a way that they are configured to form the system (for example as required).The training and / or operation of the neural network according to the invention is particularly advantageous because, in particular under the circumstances described in the present document, a very high complexity results, which in particular results from the desire to achieve and / or maintain one or more target values. This would thus not be releasable by rule-based methods sufficiently or approximately as advantageously as by means of the neural network according to the invention.Forming a number of sub-tasks from a computing task may (also) comprise or be configuring and / or adapting (e.g. automated) forming of the sub-tasks. The forming (or configuring, adapting the forming) can relate to a respective, specific, specific case and / or to the specific computing task. Thus, different (possibly not known a priori) computing tasks or computing tasks of different (possibly not known a priori) types can be executed (performed) in each case advantageously.For example, the at least one neural network can be trained, retraining and / or operated, to decide, in particular to aim, achieve and / or maintain an optimum with respect to:a number N of sub-tasks from a computing task R;on the number M and / or features of the arithmetic units used;distributing the sub-tasks among the computing units;This can be done and / or varied, adapted (if appropriate dynamically) during the course of the method in such a way that one or more target values are reached and / or held as well as possible.The neural network is preferably set up, trained, trained, retraining and / or operated for the specific or specific features corresponding to the computing tasks, type and / or number of computing units, users of the devices comprising the computing units. This enables a result optimized to a high degree to be achieved.The method preferably further comprises: providing a computation result of the one or more parts of the computation task by the one or more computation units (e.g. to a backend) and / or receiving a (des) computation result outside the one or more computation units (e.g. by a, in particular the same or networked, backend).According to a second aspect, a (second) method for executing a computing task on a plurality of computing units (in particular also to be understood as being by means of a plurality of computing units) is provided. The (second) method comprises: training, in particular retraining, a neural network to form a plurality of sub-tasks from a computing task, in particular to divide the computing task into a plurality of sub-tasks, in such a way that one or more (of the formed) sub-tasks can be executed on a plurality of computing units.Alternatively or additionally, the (second) method can comprise: training, in particular retraining, a neural network to distribute, in particular dynamically, a plurality of sub-tasks formed from a computing task to a plurality of computing units depending on an output layer of a neural network. This neural network can be one or more features described herein, in particular the (same) neural network with which the calculation tasks are formed. In this case, the neural network can be trained, in particular trained at night, to make the one or more sub-tasks executable on a plurality of computing units, in particular taking into account (e.g. for the best possible fulfilment) one or more target values (e.g. described in the present document), in particular target values relating to a superordinate computing task.The features of the (first) method according to the first aspect and features of the (second) method according to the second aspect can be combined with one another. In particular, a resulting method comprises the features of the first method and features of the second method. Therefore, the features described below may relate to the first method and / or the second method. The term method used below is to be understood in particular as the first method, as the second method or as an (advantageous) combination of the methods.The new-tonal network described in the present document is preferably a neural network which is set up and / or trained, in particular correspondingly trained, for the execution of a computing task on a plurality of computing units (according to the features described in the present document). In this case, it can be a neural network (particularly, specifically, predominantly) set up and / or trained, in particular correspondingly retrained or retrainable, for forming a number of sub-tasks from a provided computing task and / or distributing the sub-tasks to a plurality of computing units.The at least one neural network can be trained (in the method), in particular trained at night, and / or be operated to distribute and / or execute a computing task in the form of one or more parts on a plurality of computing units, in particular taking into account a target value corresponding to specific or specific features.This can be done for achieving a result of the calculation task, reducing the cost, accelerating the execution of the calculation task, etc. The achievement of a result of the computing task can comprise or be a contribution to the (overall) computing task, in particular the result of the computing task (e.g. ascertaining or the data, in particular data representing the result of the computing task).In this case, the input layer can be controlled as a function of the target value, in particular of one or more target values corresponding to the features described in the present document (in particular to be understood as data which characterize the target value).For example, the at least one neural network can be trained or trained overnight to determine one or more parts of a computing task (e.g. those which are initiated at specific vehicles or for respective computing units or per computing unit) in such a way that the execution of the computing task is optimized to a plurality of computing units from a number of computing units; and / or a result of the execution of the computing task.The term "optimizing" or "optimum" is to be understood within the scope of the present document as meaning, in particular, an (expedient) tendency of an optimum or an approach to a (theoretical) optimum and / or the holding of an optimized (improved) value, in particular apart from a certain tolerance value or approach. In this case, an optimization (e.g. in comparison with a conventional or rule-based algorithm and / or in addition to or after execution of a rule-based algorithm.The method can furthermore comprise: receiving a plurality of results of the execution of the one or more parts of the one (or in each case) a plurality of computing tasks (e.g. individually or from a plurality of computing units) and providing the result (e.g. a total result combined from partial results) for the one (or a plurality of computing tasks. The one or more computing tasks can be read in, in particular, from a backend (outside the computing units mentioned) or ordered or provided by the latter. The result of the execution of the one or more computing tasks can be provided to a backend, in particular the backend (the same backend or in a backend networked to the backend).For example, the neural network can be trained, in particular trained at night, to select one or more computing units for executing one or more parts of the computing task from a number of vehicles; and / or to assign one or more parts of the (respective) computing task to one or more computing units. It can be trained for this process to be optimized according to the determined or determinable target value.The determination of one or more parts of a computing task can comprise or be a selection of one or more parts and / or a determination of the parts of the computing task and / or a division of the (one or more) computing tasks into sub-tasks (in particular also to be understood as sub-tasks).Alternatively or additionally, the determination of one or more parts of a computing task can be carried out in such a way that (advantageous or relevant for achieving or holding a specific target value) distribution, in particular redistribution of the parts of the computing task to a plurality of computing units or between a plurality of computing units, comprises or is.For example, the determination of one or more parts of a computing task can be carried out in such a way that a number N (e.g. more than three two) of the computing tasks is expediently distributed to a number M (e.g. more than 10, 100, 1000) of the (selected) computing tasks, in particular a number M' (e.g. more than 20, 200, 2000) of the (selected) computing units (from a number of devices comprising the computing tasks).One or more steps of the method may be performed depending on the one (or more) neural networks. In particular, this can comprise or be: controlling the execution of the respective step of the method as a function of the operation (also: operation) of the one or more neural networks, in particular as a function of the output layer, in particular of the data of the one or more neural networks output or generated at the output layer.The target value can be a predetermined target value that can be set or controlled more specifically (e.g. case-specific) and / or variably and / or, in particular dynamically and / or during the execution of the method, for specific computing units, users, conditions, the execution of the method, in particular one or more operating cycles (e.g. specific operating cycles) of the method. For example, the target value may vary or be controlled in the method depending on the vehicle computing task. For example, one or more target values for a specific (e.g. current) operating cycle, per computing task (e.g. overall result of the computing task, per computing power and / or can be predefined or are applicable as a target for the average to be achieved.For example, the target value (e.g. by one or more users or operators of the method) can be adjusted, in particular controlled, in particular for one or more specific operating cycles, computing tasks and / or dynamically and / or during the execution of the method. This results in several advantages that are easily understood by the skilled person. For example, a plurality of target values (e.g. by the orderer of one or more respective computing tasks and / or by one or more users of the respective computing units or devices comprising them and / or by an operator of the method) can be taken into account. For example, the at least one neural network is trained and / or operated to aim or maintain an optimum between a plurality of target values.Alternatively or additionally, the target value can be adjusted in the method, in particular depending on a (detected) feature or the type of the computing task. The adaptation of the target value can be carried out according to a specific or predetermined mathematical relationship (e.g. parameter, formula, characteristic curve, coefficients of a predetermined formula) and / or as a function of predetermined criteria.One or more of the steps of the method can be carried out (e.g. with respect to a plurality of computing units) in such a way that a target value (which applies and / or is common overall and / or which applies to one or more computing tasks, users, computing units) is sought, in particular is reached and / or held (with an expedient approximation). This can be effected by the corresponding training and / or operation of the at least one neural network.The computing resources of the respective computing units described herein can relate (individually or according to a predefined formula) to: the CPU load (relating to general purpose unit); and / or GPU load (relating to graphics processing unit); and / or the RAM (working memory); and / or the nonvolatile memory (e.g. flash memory, EEPROM); and / or data transmission, in particular data transmission capacity, data transmission rate, data packet size, priority, response time or a delay in the data transmission. In this case, the execution of one or more steps of the method can take place depending on available resources of the one or more of the aforementioned types and / or differently for different types of resources.In particular, the execution of the computing task on one or more computing units is also to be understood as: computing tasks which are to be executed by means of the one or more computing units (at least predominantly with means, resources of the vehicles) from a number of computing units.For example, the at least one neural network is trained, in particular trained at night, to determine the one or more parts of one (or in each case a plurality of) computing tasks in such a way that the execution of the one or more parts of one (in each case) computing tasks and / or the result of the execution of the one or more parts of one (or in each case a plurality of) computing tasks (e.g. according to a predefinable target value) is optimized.For example, this takes place in such a way that a specific and / or (respectively determinable, settable) and / or automatically adaptable target value (in particular also to be understood as a weighted ratio of a plurality of target values) is sought, in particular reached and / or maintained.The term "neural network" is to be understood in particular as a deep neural network. Alternatively or additionally, the term neural network is also to be understood as meaning another form of the artificial intelligence (in particular comprising a neural network or a comparable structure). In particular, training and / or operating the one or more neural networks is to be understood as meaning an analogous application of other types of machine learning.For example, the at least one neural network, the training and / or operating thereof can comprise or be an appropriate application of one or more circumstances listed below or combinations thereof: wholly or partly supervised learning, wholly or partly unsupervised learning, at least partly reinforcement learning, multi-task learning, classification and regression trees, support vector machines, logical and relational learning, probabilistic graphic models, rule learning, instance based learning, latent representation, bio-aspirated approach.Particularly preferably, a fully or partially reinforced learning can be used. For example, the reinforcement may be dependent on a result of a previous (e.g., on one or more previous operating cycles, days, months, etc.) Execution of the computing tasks (e.g. depending on individual and / or statistical and / or cumulative results) and / or the extent to which one or more target values (e.g. described in the present document) are reached or executed, respectively.The neural networks or artificial intelligence described in the present document (which are based on the one or more models, in particular generative pre-training transformers or the functionalities) is preferably a so-called weak and / or (comparatively closely) specialized artificial intelligence.These (or the corresponding models or neural networks can be specialized for the execution of the (single or multiple) functionalities described in the present document and / or optimized to a certain extent for the execution of these and / or limited to these (at least predominantly or essentially). As a result, the method can be realized substantially more easily and / or with limited resources (e.g. using the means of the neural network) and / or can be better controlled (in practical handling).For example, this can reduce possible risks in connection with the use (e.g. a relatively universal or strong) artificial intelligence and / or make the result of the method more predictable.The term "training" is to be understood in particular as machine learning, non-rule-based learning and / or deep learning or the execution of machine learning, non-rule-based learning and / or deep learning. The training can comprise or be a (at least partially, completely or predominantly) monitored or (at least partially, completely or predominantly) unsupervised training. The training of the neural network is to be understood in particular (also) as setting up the neural network.The training (or setting up) of the at least one neural network can comprise or be an end-to-end learning (also to be considered: end-to-end learning) and / or a reinforcement learning (also to be considered: reinforcement learning or reinforcement learning). In this case, a plurality of, in particular a predominant number, or (substantially) all intermediate steps necessary for the request for one or more targets can be mapped or integrated in the (at least partially trained) neural network.The training of the neural network is to be understood in particular (also) as setting up the neural network.The operation of the neural network is or comprises in particular a control of the neural network, in particular of the input layer (one or more inputs) of the neural network. Furthermore, a change in the operating mode of the at least one neural network can likewise be counted for operating the neural network.In particular, the at least one neural network is trained and / or operated depending on a combination of at least two parameters (corresponding to one another) (e.g. data of two different types or characteristically different size described in the present document). In this case, one or more multidimensional optima (e.g. for executing the tasks other than the vehicle according to two or more than two different target values) can be sought, in particular can be achieved or held with a certain approximation.The training and / or operation of the one or more neural networks can take place according to one or more predetermined targets or target values (described, for example, within the scope of the present document) that can be determined, adjusted and / or dynamically changed (for example, by a user and / or dispatcher) in the method.The term "target value" can also always be understood within the scope of this document to mean a target value range, for example a value range restricted upward and / or downward and / or standing as a function of a further variable (for example a parameter). The target value can also be understood as an attempt to achieve an (expedient) optimum of the respective dimension. Such an optimum is in particular a local or global minimum or maximum of the respective measure (in particular to be understood as a suitable approximation to the minimum or maximum). The reaching and / or holding of the one or more target values can therefore also be understood as an optimization. The at least one neural network can therefore be trained and / or operated, a plurality of target values, an overall optimum and / or an optimum for a (determined or determinable weighting of a plurality of target values are sought, in particular are reached and / or maintained.It can be one or more specific target values, in particular (for example, selectable per computing task and / or time interval), or a target value which takes into account a plurality of the target values described in the present document.The method can be carried out in such a way that, determining, in particular selecting and / or reading in, (e.g. according to a specific or determinable criterion) one or more parts of the computing task which are to be carried out on the one or more computing units; and / or forming the plurality of sub-tasks on the basis of the computing task, in particular dividing the computing task into a plurality of sub-tasks; and / or distributing the sub-tasks for carrying out a computing task on one or more computing units; is carried out as a function of a target value, in particular a dynamically predefinable target value.The one or more target values can preferably be taken into account during training, in particular retraining of the at least one neural network and / or during operation of the at least one neural network. In this case, the at least one neural network can be operated, in particular controlled (e.g. at its input layer), as a function of the one or more target values.For example, the neural network can be controlled as a function of a specification, in particular a prompt, (e.g. by an operator of the method and / or user of the devices comprising the computing units, in particular dynamically determinable specification).The training, in particular retraining, of the one or more neural networks can be carried out, in particular in a plurality of phases, over a long period of time or over a plurality of executions or operating cycles of the method. The training of the at least one neural network can take place at least partially, in particular only or predominantly automatically, during the operation of an already at least partially trained neural network or in a separate process. The method can comprise a change, in particular a longer-term, slow update, of the neural network comprising a plurality of the operating cycles, depending on a plurality of the first data sections.The training (in particular also to be understood as the process of training) can comprise or be:determining training data, in particular on the basis of the data from previous executions of one or more (e.g. corresponding to similar features) computing tasks; and / ordetermining and / or causing the backpropagation, in particular on the basis of the data from previous executions of one or more (e.g. corresponding to similar features) computing tasks; and / orselecting or adapting one or more parameters of the training process, in particular on the basis of the data from previous executions of one or more (e.g. corresponding to similar features) computing tasks.One or more aspects of training described above may be performed and / or adapted depending on the (e.g. processed and / or filtered) data from previous executions of computing tasks.The term "operating" of the at least one neural network can be understood to mean using the at least one neural network, in particular also as the neural network, as a function of or on the basis of the data (learning data) of the neural network (e.g. the neural network trained according to one or more features described in the present document), in particular in the active operation.The operation of the at least one neural network can alternatively or additionally comprise or be the determination of one or more parts of one or more vehicle-external computing tasks as a function of the output layer of a neural network; and / or the initiation of the execution of one or more parts of one or more computing tasks on one or more vehicles from a number of computing units by means of the neural network.The at least one neural network can be controlled during its operation (at its input layer) and / or (directly or indirectly, e.g. including a further data processing step) control the determination of one or more parts of one or more computing tasks.Depending on the variant of the method, the method comprises or is, for example, the training (pretraining or retraining) of the neural network. In this case, it is to be understood as "the neural network is trained (pretraining or retraining)" and / or the operation of the already at least partially trained neural network (it may be a second method). In this case, it is to be understood as "the neural network is already trained (e.g. at least to an appropriate degree) (pretraining or retraining respectively night-trained)". The method can therefore comprise or be the first method and / or the second method.For example, the at least one neural network (e.g. to be used during operation) can be understood as features of or data on the basis of the at least one (e.g. at least partially trained) neural network. For example, the data are generated during training of the at least one neural network on any computing unit, in particular the computing unit of the backend, in order to then make them operable on one or more other computing units, in particular in (e.g. respective, specific) vehicles, of a computing unit that is central or networked. The operation of the at least one neural network can comprise or be a setup of the at least one neural network as a function of the data of in each case at least one corresponding neural network or neural network trained for the corresponding use.Each of the information described in the present document may comprise or be a measure of probability (corresponding to the corresponding parameters or with respect thereto). For example, the computing power of one or more computing units that is at least likely to be available may be a measure of the probability or may comprise a measure of the probability (e.g. in addition to a quantitative and / or qualitative measure of the computing power).In an example which can be combined with all the features described in the present document, the determination of a plurality of parts of the computing task, in particular of part tasks of the at least one computing task, which can each be executed on one of a plurality of computing units, can be carried out as a function of one or more feedbacks of the one or more devices comprising the computing units with respect to a progress in the execution of one or more parts of the computing task and / or already executed parts of the computing task. For example, the neural network can be operated, in particular controlled, as a function of feedback.The computing task preferably comprises a program code comprising one or more instructions (also to be understood as an executable program, a binary code, etc.) and / or (from the outside) read-in (to be processed) data, in particular data, which are to be processed by means of said program code. For example, the computing task can comprise a program code to be transmitted to the (respective) computing unit once or comparatively rarely and (e.g. always new) (external) data (e.g. data packets) to be transmitted to the respective computing unit several times or comparatively frequently.The computing task is preferably a computing task that requires considerable computing power. For example, the computational task may have a scope that requires processing by a processor (e.g., laptop processor) (e.g., typical or commercially available at the time of application of the method described herein) over several minutes, hours, days, weeks, or months.The computing task, in particular the program code and / or data of the (external) computing task, can be independent of a (different) functionality of the (respective) devices comprising the respective computing units, in particular independent of all essential or any of the (respective) devices.For example, the computing task (e.g. provided from outside the respective device comprising the computing unit) can (also) be independent of (essential or all) system functionalities and / or user functionalities of the respective devices (comprising the computing units). Alternatively or additionally, the computing task can at least partially comprise program code and / or data which are independent and / or separate from one, in particular from any, program code or data of the devices comprising the computing units.For example, the system may be configured to perform various (virtually any) computing tasks that are not known a-prior, in particular within certain limit values (e.g. with respect to performance), in particular computing tasks of any types that are not known a-prior, on a plurality of (optionally different or different types) computing units, in particular taking into account the respective target values.The computing task, the overall computing task and / or at least one sub-task can be determined or configured for a purpose which is completely or substantially independent (at least predominantly or substantially) of the device, data and / or functionalities of the devices comprising the computing unit.For example, in an embodiment of the method, no substantial or permanent change of the functionality and / or data of the device has to take place. These can also serve (optionally further and / or if necessary) further purposes, in particular their purposes.For example, the system can be configured to prevent or prevent interaction or mixing of the data and / or the executable program code or the instructions of the computing task with or any data or programode that is otherwise customary for the computing unit or belongs to the device comprising the respective computing unit and is used in particular for executing a system functionality or a user functionality.The method or system described in this document (e.g. each of the variants of the system) can thus be configured to use the computing resources of at least the plurality of computing units, in particular a number or group of computing units (corresponding to specific or specific features) or devices comprising the computing units) in a particularly efficient manner. In this case, a plurality of computing units, in particular belonging to a number or group, can advantageously be used to carry out (also to be understood as to be able to be executed or solved) an external or independent, external computing task (not known a priori, specified from the outside or predefinable with a high degree of freedom of design). In this case, the benefit and / or the ecological balance of the parts or computing resources (serving other purposes, for example) of the computing units (in particular of the one or more hardware components) can be increased, in particular in an efficient manner.The method may further comprise:Executing the (at least one, respective) part of the computing task or partial assignment, in particular in addition to a program code of a computing task that is specific or specific to the respective computing unit, at least partially a new program code, in particular for operating and / or optimizing the respective computing unit and / or respective computing component of the computing unit, is applied to the computing unit, in particular to one or more computing components of the respective computing unit, in particular in the form of firmware.In this case, the new program code, in particular the new program code for operating and / or optimizing the computing unit computing component, can differ from and / or be independent of the program code, in particular from the program code for operating another dedicated functionality specific to this and / or from the program code which is executable or executed by the computing component during operation of the device (typically serving a different purpose) comprising it.The method can comprise applying at least partially new program code, in particular in the form of firmware, to the one or more of the plurality of computing units, in particular one or more computing components of the computing unit) for executing the one or more sub-tasks. In this case, the program code (alternatively or additionally to instructions to be applied to the data) can comprise parts of an operating system for operating the one or more computing components.For example, firmware components (each adapted to the execution of the specific computing task or computing task of a specific type, in particular software component optimized to it, in particular comprising parts of a (adapted) operating system, can be transmitted to the respective computing unit or computing component and / or selected and / or activated on it. This can serve for improved, in particular optimized, effectiveness of the processing of the one or more computing tasks, in particular of a specific (e.g. expected, jointly transmitted) computing task or a computing task of a specific type.Preferably, at least a portion of a program code (otherwise, normally, typically) used on the (respective) computing unit (e.g. serving for executing a functionality of the respective device) and / or data of the computing component for executing the (external, non-typical) computing task can be deleted at least temporarily and / or copied or moved into another memory area inside or outside the computing component.The method can comprise placing one or more computing components of one or more computing units in a first computing mode for executing the (respective) part of the computing task.The (at least one) neural network can preferably be trained and / or operated to carry out a dynamic and / or flexible distribution of the sub-tasks to a plurality of computing units and / or to a plurality of (specific) computing components of the (respective) computing units (e.g. also within the computing units). For example, in such variants of the invention, the individual computing components (possibly belonging to a higher-order computing unit) are to be understood as computing units. In this case, these may be the computing units in the sense of the present document.As a result, (at least cumulatively) a substantially improved utilization, in particular the degree of utilization, of the computing units available in each case for executing at least part of a computing task and / or (in each case available) computing resources of the computing units can be enabled.According to a further embodiment, distributing the one or more sub-tasks to a plurality of computing units comprises or is a control of the execution, in particular a dynamic and / or adaptive control of the execution of the one or more respective sub-tasks.The control and / or distribution can take place during the execution of the (respective) one or more sub-tasks.For example, the control can be carried out in such a way that suitable points in time between an at least anticipated processing of a plurality of the one or more sub-tasks are sought, in particular observed.For example, the execution of threads of the one or (respective) multiple sub-tasks can be controlled individually, selectively and / or a scheduling and / or at least partially parallel (also to be understood as: quasi-parallel) temporally overlapping and / or at least partially sequential execution of threads of the sub-task (of the computing task) can be controlled (e.g. caused, intensified, throttled, changed.For example, forming one or more sub-tasks may include or be forming threads (e.g., within the computing task or within at least partially separate sub-tasks). In particular, the distribution of sub-tasks can include or be the distribution of threads (e.g. to a plurality of computing units or computing components of a computing unit).For example, the degree of parallelism, distribution of threads to pipes and / or processor cores of the respective sub-tasks can be controlled, in particular dynamically and / or during the operating cycle or during their execution. This aspect of the invention can also be applied to (entire, individual, autonomous) computing tasks.The adaptively controlling refers in particular to the tendency to reach and / or hold one or more, in particular dynamically variable target values.In particular, the control (e.g. to a plurality of the sub-tasks on (in each case) a plurality of computing units) is carried out in such a way (e.g. the balance between the intensity of the control is selected in such a way that the aim is to achieve or maintain the target values as well as possible.According to a further embodiment, the control of the execution of the one or more sub-tasks is or comprises: a control of an interface provided in the one or more sub-tasks, in particular by means of the interfaces provided in the executable code of the respective sub-tasks; and / or a control of the operating system of the respective computing units, wherein the operating system influences the execution of the respective sub-tasks depending on the control.The controlling may (also) comprise or be a configuring and / or adjusting. The control (or configuring, adaptation) can relate to a specific case, in particular to the specific computing task and / or (a plurality of) specific computing units. Thus, different computing tasks or computing tasks of different types (possibly not known a priori) can be executed (performed) in each case in an improved manner.According to a further embodiment, distributing one or more sub-tasks on a plurality of computing units comprises or is redistributing one or more of the sub-tasks, in particular during handling of the computing task and / or during executing one or more of the sub-tasks.The redistribution can preferably comprise or be: a stop, a termination of the execution of one or more sub-tasks (e.g. before the sub-result is present), a removal of a part of the sub-task that has not yet been executed or a renewed formation of a sub-task from the sub-task; and / or transmission of the sub-result and / or the new sub-task to the backend or to another of the plurality of computing units. Alternatively or additionally, the redistribution may comprise aborting execution of one or more sub-tasks from the plurality of sub-tasks. One or more of the aforementioned method steps can be carried out depending on the output layer of the at least one neural network.According to a further embodiment, the control of the execution of one (or more) first sub-tasks is carried out on a first computing unit:depending on the control of the execution of one or more second sub-tasks on one or more second computing units (e.g. depending on the corresponding control signals or control carried out); and / ordepending on the progress, intermediate result and / or result of executing one or more second sub-tasks on one or more second computing units.In this case, the at least one neural network can be operated on the basis of the data in a characterizing manner to control the execution of one or more second sub-tasks on one or more second arithmetic units; and / or in a characterizing manner to progress the execution of one or more second sub-tasks on one or more second arithmetic units. For example, the at least one neural network is trained, in particular retrained and / or operated, in particular controlled at the input layer, on the basis of these data.In this case, the at least one neural network can be trained, retrained and / or operated, the role of first arithmetic units and / or second arithmetic units can be assigned to the arithmetic units, in particular dynamically, and / or the dependence between the aforementioned, in particular dynamic, processes can be modified.According to a further embodiment, the at least one neural network comprises:a generative pre-trained transformer (GPT); and / ora generative, in particular statistical model; and / ora statistical model trained at specific computing units, users of the devices comprising the computing units, computing tasks;or is constructed on the basis of such a system and / or controlled by such a system.According to a further embodiment, the at least one neural network is or is trained, in particular retrained, to aim, in particular at least presumably to reach and / or maintain, a target value with respect to the computing task.For example, the method may include:training, in particular retraining, a neural network from a computing task to form a plurality of sub-tasks, in particular to divide the computing task into a plurality of sub-tasks; and / orforming the plurality of sub-tasks on the basis of the vehicle-external computing task;depending on the at least one target values, in particular depending on a plurality of target values (e.g. related to different criteria).The target value (e.g., each of the plurality of target values) may relate to the execution of a plurality of, in particular all, essential parts of the computing task, in particular all essential sub-tasks, on one or more vehicles. The one or more target values may be cumulative and / or conserved with respect to a sub-task.For example, the at least one neural network is trained, retraining and / or operated to aim for, in particular (apart from a certain tolerance value), an optimum between the reaching and / or different target values, in particular taking into account a (predetermined) weighting of the target values that can be predetermined (e.g. by one or more users).For example, a weighting, in particular as a respective importance of the target values, e.g. speed, compliance with a deadline, costs, an ecological balance value, reliability of the execution of a computing task, can be predefined or can be predefined (e.g. by users, operators, ordering agents of the computing task). The at least one neural network can then receivea number N of sub-tasks from a computing task R; and / oron the number M and / or features of the arithmetic units used; and / ordistributing the sub-tasks among the computing units; and / or selecting, changing them in this way; and / orcontrolling execution of the one or more sub-tasks of the computing task;This results in the best possible overall result of the execution of the method, in particular of the specific operating cycle of the method.According to a further embodiment, the training, in particular night training and / or operation (e.g. control, prompt in current operation) of the at least one neural network is carried out as a function of one or more parameters with respect to the one or more computing units.In this case, the at least one neural network can be trained, in particular retrained, as a function of the data (e.g. from different time intervals or operating cycles from the past), wherein the data relate to: the parameters of the sub-tasks formed (in the past) of one or more computing tasks; and / or computing units selected from a number of computing units (e.g. computing units in total question) (in the past); and / or assignments of one or more sub-tasks to one or more computing units; and / or from the reaching and / or holding (e.g. from a measure of the reaching or holding) of one or more target values with respect to the (respective) computing task.According to a further embodiment, the at least one neural network is or is trained, in particular retrained, depending on the computing power available and / or used in the past of one or more of the plurality of computing units or of a superordinate number of computing units.In this case, the at least one neural network can be trained, retraining and / or operated to predict the (respectively) available and / or (for specific, respective) sub-tasks required computing power. The result of the prediction can be taken into account in the method. The prediction can relate to one or more time intervals. For example, a prediction can relate, for example, overall to several seconds, minutes or hours. The time intervals can be small or very small in comparison with the total (prospective) duration of the execution of the computing task or the sub-tasks (e.g. a few milliseconds, a few seconds or a few minutes).For example, depending on the (in this way) trained, in particular night-trained, neural network: forming a number of sub-tasks from one computing task; and / or distributing sub-tasks from the number of sub-tasks to a plurality of computing units; and / or controlling the execution of a first sub-task on the one or more computing units; may be executed. This can be effected, for example, depending on (explicit or implicit) results of the (respective) predictions (for example for a plurality of the arithmetic units).Alternatively or additionally, it can be an end-to-end training and / or operation of the at least one neural network (also) depending on available and / or used computing power of one or more of the plurality of computing units, in which the formation of a number of sub-tasks from one computing task; and / or distribution of sub-tasks from the number of sub-tasks to a plurality of computing units; and / or control of the execution of a first sub-tasks on the one or more computing units takes into account the respective expected available and / or resources required (by the sub-tasks) (implicitly, by end-to-end training and / or operation already).For example, an anticipated future change in the respectively available computing power or corresponding predictive values can already be taken into account when forming a number of sub-tasks from one computing task, distributing, in particular redistributing, sub-tasks to a plurality of computing units and / or when controlling the execution of the one or more sub-tasks to (respective) computing units. In this case, the prediction may include or be one without explicit output of the predicted parameters.According to a further embodiment, the at least one target value relates to a cost required for executing at least part of the computing task, in particular a number of sub-tasks (e.g. the target value indicates the meaning of the cost and / or a value, in particular in units of value, in particular in units of money).This can be the following costs:electricity costs for executing at least part of the computing task, in particular to be understood as all essential parts of the computing task; and / ordata transmission costs for transmitting at least a part of the computing task, in particular a number of the plurality of sub-tasks, to one or more computing units; and / ordata transmission costs for transmitting at least a part of the one or more calculation results, in particular all essential calculation results, from one or more calculation units, in particular within the scope of providing the overall result of the calculation task.In particular, the at least one neural network can be trained, in particular trained at night; and / or operated; the formation of a number of sub-tasks from one computing task; and / or the distribution of sub-tasks from the number of sub-tasks to a plurality of computing units; and / or the control of the execution of a first sub-task on the one or more computing units can be carried out in such a way that the cost is optimized or a target value relating to the cost (e.g. per computing task, in particular as a total result, in the current operating cycle or on average) is sought, reached and / or maintained.The term "overall result" is to be understood in particular as meaning the result requiring a substantial, predominant part of the computing power.According to a further embodiment, the method comprises: making, in particular to be understood as a negotiation of an agreement with respect to an compensation and / or a counter-performance for the execution of the computing task on one or more of the plurality of computing units or the devices comprising the computing units; and taking into account the made agreement in the method, in particular when: forming a number of sub-tasks from a computing task; and / or distributing sub-tasks from the number of sub-tasks to a plurality of computing units; and / or controlling the execution of a first sub-task on the one or more computing units; (also) depending on the made agreement.For example, a calculation, accounting, compensation of costs and / or bids of one or more vehicles can be carried out by means of a micropayment (in particular also to be understood as: micro transactions, millipayment, nanopayment or picopayment) or by means of a data processing linked to a micropayment and / or configured according to one or more features of the micropayment. For example, the MPTP (=Micro Payment Transfer Protocol) can be used in the method.Preferably, compensation of advantages and / or disadvantages (alternatively or in addition to a currency or money units) which have been produced or expected on account of the execution of one or more steps of the method can be effected in specific (e.g. exchangeable or convertible) units. The units can comprise or be value units, e.g. money units.Alternatively or additionally, the units may identify an ecological characteristic, e.g. an at least prospective ecological balance (e.g. with respect to the result of at least a part of the computing task and / or per computing power), discounts and / or rights. For example, a calculation, accounting, compensation can be carried out on the basis of a vehicle or the number of vehicles.The method can be operated (in particular also to be understood as: controlled or executed) by a negotiation, optimization, accounting, accounting, compensation of units (e.g. controlled by a neural network, in particular the neural benefit) dividing the computing task into a plurality of sub-tasks and / or forming a group of vehicles for executing a computing task by means of (expediently set up, in particular trained) bots.The bots can be configured to determine and / or agree (e.g. by means of an operative connection to one another or a message exchange) a solution which is respectively improved according to predefined target values.According to a further embodiment, the meeting, in particular negotiation of the agreement with respect to a reward and / or a counter-performance for the execution of the computing task takes place on one or more of the plurality of computing units or the devices comprising the computing units by means of a neural network, in particular the neural network or a bot operated by the neural network.In this case, the neural network can be trained and / or operated to minimize the compensation and / or the counter power for the execution of the computing task on one or more of the plurality of computing units; wherein the training takes place or took place in particular as a function of individual compensations and / or counter powers for computing units and / or the entire compensation and / or counter power for one or more computing tasks in the past; and / or the operation takes place in particular as a function of individual compensations and / or counter powers for computing units.According to a further embodiment, the method comprises executing one or more sub-tasks on one or more computing units, by means of a further computer program which is used for operating and / or optimizing one or more computing components of the one or more computing units, for executing the one or respective ones of the plurality of sub-tasks, wherein the computer program, in particular in the form of firmware, is applied and / or operated on the one or in each case a plurality of computing units.In this case, the application of the computer program can (also) be carried out as a function of the operation of a neural network, in particular of the neural network or of a further neural network, in accordance with one or more features described in the present document.According to a further embodiment, the method comprises: receiving a plurality of results from the plurality of sub-tasks, and ascertaining an overall result for the computing task as a function of the plurality of results; and: providing the overall result of the computing task.The provision can be effected to an (external) device. For example, a device networked to the device, from which the computing task has been provided. The overall result is to be understood in particular as a result of the execution which comprises the (partial) results of most or all of the essential parts of the computing task and / or which contains the essential part of the executed computing power.According to a further embodiment, it is included or isthe computing task is a computing task from a vehicle, and the one or more computing units are computing units outside the vehicle; orat least one computing task comprises or is a vehicle-external computing task, and the one or more computing units are computing units of one or more vehicles; orAlternatively, the invention comprises a variant in which the said plurality of computing units are different computing units of the same vehicle.According to a further aspect of the invention, a neural network is described. The neural network is configured, in particular trained and / or retraining or is retraining, to form a plurality of sub-tasks from a computing task, in particular to divide the computing task (R) into a plurality of sub-tasks. In particular, such that one or more sub-tasks can be executed on a plurality of computing units (for achieving the result, in particular an overall result of the computing task).Furthermore, the neural network can be configured, trained or retrainable for (directly or indirectly) causing, controlling the execution and / or relocation of a computing task on a plurality of computing units from a number of computing units (e.g. corresponding to a specific number or number of specific features) for achieving a result of the computing task.According to a further aspect of the invention, a (first) system for executing a computing task on one or more vehicles is described. The (first) system is configured to form a number of sub-tasks from a computing task depending on the output layer of a neural network; and to distribute (e.g. in each case transmit) sub-tasks from the number of sub-tasks to a plurality of (e.g. more than 2, 10, 100, 1000, 10000, 10000) computing units from a number of computing units for achieving a result of the computing task.The (first) system is in particular further configured to carry out the method according to one or more features described in the present document. The (first) system can comprise a plurality of computing units (e.g. as parts of specific devices) and at least one external device, in particular a backend.According to a further aspect of the invention, a (second) system for executing a computing task on one or more vehicles is described. The (second) system is configured to train, in particular train night, a neural network to form a plurality of sub-tasks from a computing task, in particular to divide the computing task into a plurality of sub-tasks, in such a way that one or more sub-tasks can be executed on a plurality of computing units.The (second) system is in particular further configured to carry out the method according to one or more features described in the present document. The (second) system can comprise a plurality of computing units (e.g. as parts of specific devices) and at least one external device, in particular a backend.For example, the system can be configured to set up, train and / or operate a foundation model, wherein the foundation model takes into account the statistical dependencies in connection with: the use of computing units, in particular relating to a specific number of computing units corresponding to a specific feature, in particular by specific users; and / or with the use of computing resources and / or of the available computing resources of one or more computing units, e.g. from a specific number of computing units or a number of computing units corresponding to a specific feature.Alternatively or additionally, the system can be configured to read in, operate and / or train night such a foundation model (e.g. from a further or external source). This can involve retraining with respect to the data, which are at least predominantly relevant or representing for specific computing tasks and / or computing units, users of the computing units or the devices comprising the (respective) computing units, and / or the conditions prevailing during use, and / or a measure of reaching or holding one or more target values with respect to the computing task.The at least one neural network may be configured, trained, and / or operated to achieve a measure of achieving or holding one or more target values (e.g., reducing costs, a faster and / or more reliable result of the computing task. This can be effected, for example, for specific or specific features corresponding computing tasks, computing units, in particular computing components of the computing units, users of the computing units, conditions during the use of the computing units (e.g. characterized by other purposes).The system may comprise the first system and / or the second system, in particular any combination of the systems described herein. In this sense, these can identify features of the (overall) system and / or form (partial) systems of the system.According to a further aspect of the invention, a computer program, in particular a computer program product comprising a computer program, is described, wherein the computer program is designed, when it is executed on at least one arithmetic unit, to execute the method described in this document.The computer program can comprise, for example, the program code with which one or more vehicles are set up (during their production or later, in particular already after the delivery of the vehicle to the user) for executing the method. This computer program differs in particular from the program code (which can be used temporarily or depending on a current demand) of the (external) computing tasks.For example, the computer program is configured (e.g. by means of the new program code described in this document, which can differ from the program of the computing task) for a, in particular temporary or temporary, executable, modification of an operating system or of a part of a basic software of the at least one computing component of one or more computing units (also to be understood as: the specific computing unit), in particular for operating and / or optimizing the computing component for computing tasks corresponding to one or more specific or specific features (e.g. limit values, patterns, purposes).For example, the computer program product can be provided to the (respective) computing unit, with which the computing unit, in particular one or more computing components of the computing unit, is set up, in particular adapted and / or optimized, to execute one or more variants of the method.The computer program product can be designed as an update of a previous computer program, which comprises the parts of the computer program or of the corresponding program code, for example, in the context of a function extension, for example, in the context of a so-called "remote software update". The computer program product comprises, in particular, a medium readable by the data processing device, on which the program code is stored, or at least one encrypted file. According to a further aspect, a program product is described which comprises an authorized access right to stored data of the computer program product.It should be noted that the methods, systems, and systems described herein may be used both (each) alone and in combination with other methods, devices, and systems described herein. Furthermore, any aspects of the methods, systems described in this document can be combined with one another in a variety of ways. In particular, the features of the claims can be combined with one another in many ways.The invention is described in more detail below on the basis of exemplary embodiments without limiting generality. The following are shown: FIG. 1 illustrates an example of the system S including a backend BE and some of the plurality of computing units 11, 12, 13, 14... The arithmetic units 11, 12, 13, 14,... have or are hardware.The computing units 11, 12, 13, 14, 15... each comprise e.g. one or more processors, in particular (GPP) and / or specialized or highly parallel processors, e.g. graphics processors (GPU) and / or volatile memory (RAM) and / or nonvolatile memory (EEPROM), etc.At least two of the plurality of computing units can each be computing units of different types.The arithmetic units 11, 12, 13, 14,... are:different in respect of their respective resources of different types and / or performance. In particular, these may include different types of hardware; and / orspecializing or optimizing to calculation of (various, further) computing tasks (which may differ significantly from the external computing task with regard to their type); and / orused or operated by different (respective) users; and / orused differently (e.g. over time) and / or loaded differently (by performing other computing tasks). In particular, it may be utilization and / or loading that correlates with one another in a significant manner and / or a pattern (possibly not explicitly known); and / orthe utilization and / or utilization of different, in particular independent of the computing task, in particular dependent on correlation conditions with one another in a significant manner and / or pattern (possibly not explicitly known); and / ordifferent, in particular having different effective power consumptions and / or operating costs, in particular when performing different types of computing tasks.For example, the computing units 11, 12, 13 can be connected (for example, in each case at different time intervals) to different power sources (for example, with a different and / or (in each case) variable power price and / or have current energy requirements or energy requirements predicted for a time interval in the future.The (external) computing task is in particular a computing task which is to be executed (successfully) within a specific time or up to a specific deadline, in particular with a high probability. For example, an excess of the execution time and / or a specific deadline must be below the probability of 10 -2, 10-3, 10-4, 10-5, 10-6.In particular, the computing task R is a computing task R that is superordinate and / or supplemented with one or more (preferably all or all essential) of the sub-tasks R 1, R 2, R 3,... (formed in the method). In particular, this is emergized with respect to the one or more sub-tasks. This is also to be understood in particular in such a way that no (comprehensive) knowledge about the computing task R is contained in any or most of the sub-tasks.In the case of such computing tasks R, the invention, in particular the formation, distribution, control, redistribution of the sub-tasks, results in a particularly great advantage. The overall (mathematical) complexity that results would otherwise not be sufficiently resolved in practice using other known technologies.The backend BE comprises a device K which is set up to divide a computing task R into a plurality of sub-tasks R 1, R 2, R 3 to be caused to different computing units. The forming S 40 takes place depending on the output layer of the at least one neural network DNN.The backend BE further comprises a device A which is configured to combine the partial results E 1, E 2, E 3 to form an (at least preliminary) overall result E.The backend (e.g. also to be understood as a source of the external computing task) BE can comprise a neural network DNN, GPT (described in the present document). This is also to be understood in such a way that it is (remotely) controlled) by the neural network DNN, GPT (for example at least partially spatially swapped).For example, the device K used to form a computing task R and / or the device A used to combine the partial results E 1, E 2, E 3 to form an (at least preliminary) overall result E is controlled by the neural network DNN, GPT. This is also to be understood as being substantially influenced by the result of operating S 30 the neural network DNN, GPT.The individual devices, e.g. K, A, can be operated spatially together or separately from one another.The system S is configured for a dynamic and / or adaptive mode of operation (e.g. with respect to various changes in the course of the method V, V 1, V 2, in particular with respect to changes in the data processed in the method).For example, the formation of the sub-tasks R 1, R 2, R 3 and / or the distribution S 50, in particular redistribution S 50' and / or execution S 70, is adjusted dynamically, for example in a controlled manner S 60, for example depending on new ascertained or changed information, in particular taking into account the one or more target values.FIG. 2 shows an exemplary flow chart or sequence of the method according to the invention on the basis of exemplary steps S 10, S 20, S 30, S 40, S 50, S 60, S 70 of method V, V 1, V 2. The method V, V 1, V 2 can be executed cyclically, with a plurality of successive operating cycles.The solid arrows indicate a preferred proposed sequence of steps. Further, the arrows may also mean that the step pointed to by the arrow is executed depending on the result of the step at which the arrow starts. The dash-dot arrows each mean an influence during an operating cycle, in particular a control of the execution of a step pointed to by the arrow, depending on the result of one or more steps (from a preceding operating cycle) from which the arrow or the arrows originate. The method can comprise in particular only a part of the sequence shown and / or also further steps.In step S 10, the data are ascertained with respect to one or more of the (previous) executions of one or more of steps S 20, S 30, S 40, S 50, S 60, S 70 or (appropriate, expected, assumed or predicted) initial values are used. The data can (in each case) be qualitative and / or quantitative measures, in particular characterizing the result and / or the process of carrying out one or more of the steps mentioned.The data can be merged and / or processed, e.g. aggregated, from a plurality of computing units 10, 11, 12, 13,... and / or operating cycles carried out in the past (e.g. in step S 10 and / or in a further step not shown in FIG. 3 ), in particular according to a predetermined mathematical dependency.Then, depending on the data:the training data for training the one or more neural networks DNN, GPT can be determined, in particular updated; and / orthe back propagation (e.g. the data for executing the back propagation) can be determined and / or back propagation can be initiated and / or controlled; and / orselecting or adapting one or more parameters of the training process.Alternatively or additionally, the merged or aggregated and / or processed data (in step S 10) can be provided or used for operating S 30 the at least one neural network DNN, GPT.The term "data" is also understood within the scope of the present document in particular as data ascertained depending on the data or on the basis of the data.In step S 20, the training and / or retraining S 20 (e.g., according to the above descriptions) may be performed.In step S 30, the operation of the one or more neural networks DNN, GPT (e.g. as described above) may be performed.The operation S 30 of the at least one neural network DNN, GPT can (also) comprise or be a case-specific control S 30 or control S 30 relating to one or more (specific) operating cycles.The operation S 30 of the at least one neural network DNN, GPT can be in particular a (e.g. case-specific, target value-specific) prompt.For example, the input layer of the at least one (or of the plurality of corresponding) neural networks DNN, GPT is controlled, in particular prompted, depending on the data determined in step S 10.In step S40, the formation of a plurality of sub-tasks R1, R2, R3,... from the one (external, provided) computing tasks R; and / or distribution S50, in particular redistribution S50', of sub-tasks R1, R2, R3,... on or between a plurality of computing units 10, 11, 12, 13, 14, 15... This can be carried out as a function of the output layer of the (correspondingly set-up, trained S20 or retrained S20) neural network DNN, GPT.This can be done dynamically and / or taking account of the changes and / or new data and / or target values.In step S 50, (alternatively or additionally) a (further) distribution or redistribution of the sub-tasks R 1, R 2, R 3,... or sub-tasks of one or more sub-tasks R 1, R 2, R 3,... can be carried out to a plurality of computing components or sub-components of the respective computing unit 10, 11, 12, 13,...The method V, V 1, V 2 can also be executed or executed within a computing unit 10, 11, 12, 13, 14, 15,... This can (also) take place during the execution of the (current) operating cycle of the method V, V 1, V 2, e.g. during the handling of the (respective) computing task R. The control S 60 of the execution of the one or more sub-tasks R 1, R 2, R 3 can be carried out by means of the interfaces provided in the executable code of the respective sub-tasks R 1, R 2, R 3,... ; and / or by the control S 60 of the operating system of the respective computing units 10, 11, 12, 13, wherein the operating system influences the execution of the respective sub-tasks R 1, R 2, R 3,... depending on the control S 40.The control S 60 of the execution of a first sub-task R 1, R 2, R 3,... on a first computing unit is carried out: depending on the control S 40 of the execution of one or more second sub-tasks R 1, R 2, R 3,... on one or more second computing units 10, 11, 12, 13; and / or depending on the progress and / or intermediate result or result of the execution S 70 of one or more second sub-tasks R 1, R 2, R 3,... on one or more second computing units 10, 11, 12, 13.In step S 70, the one or (in each case) a plurality of parts of the computing task R is executed on one or a plurality of vehicles 10, 11, 12, 13,... The execution S 70, in particular the operation of the execution, can be controlled S 60 (also) depending on the neural network DNN, GPT; in particular dynamically and / or during operation.The control S 60 of the execution S 70 of one (or more) first sub-tasks R 1, R 2, R 3,... (e.g. on a first computing unit 10, 11, 12, 13,... ) can be controlled:depending on the control S 60 of the execution S 70 of one or more (other) second sub-tasks R 1, R 2, R 3,... (e.g. on one or more second computing units 10, 11, 12, 13; AND / ORdepending on the progress, intermediate result and / or result of executing S 70 one or more second sub-tasks R 1, R 2, R 3,... (e.g. on one or more (other) second computing units 10, 11, 12, 13.Subsequently, providing one or more results E 1, E 2, E 3,... (Calculation results) of one or more sub-tasks of the calculation task are carried out by the plurality of calculation units 10, 11, 12, 13. The provision can be made to the backend BE. A plurality of results can be received by the backend BE and / or fused outside the one or more computing units 10, 11, 12, 13, in particular in the backend BE, to form an overall result.Furthermore, the data can be determined, in particular updated and / or aggregated, as characteristic of the execution of steps S 20, S 30, S 40, S 50, S 60, S 70 (for the first time or repeatedly) in the (next) step S 10.(further) optional features or steps of the method V, V 1, V 2 are described below, which features or steps can be combined both with one another and with all further steps described in the present document.In a further example, which can be combined with all the features described in the present document, the computing task can be the handling of a blockchain or mining.Calculations necessary for handling blockchain (e.g., on multiple computing units 11, 12, 13, 14... ) can be calculated in the form of the sub-tasks. For example, at least a part of the computing task R comprises mining for generating and / or handling a cryptocurrency.For example, the computing units can be nodes of a system S for handling a blockchain or can be used as such (e.g. as required).For example, the method V, V 1, V 2 is a method for generating and / or handling a cryptocurrency.It is particularly important that a system S can be or is built up (at least temporarily), that the system S is set up, in particular optimized, to use a plurality of computing units 10, 11, 12, 13 in a dynamic and / or flexible manner (including processing large) computing tasks R or a plurality of computing tasks R.An overlapping computing task R can be distributed to a plurality of computing units 10, 11, 12, 13 and / or to a plurality of different time intervals and / or controlled (if appropriate during the execution of respective sub-tasks R 1, R 2, R 3,... of the computing task R, depending on the result of the ascertainment, in particular or prediction of a computing power, which is at least presumably available, of one or more computing units 10, 11, 12, 13.Exceptionally, the sub-tasks R 1, R 2, R 3,... of the computing task 3 can comprise or be so-called threads of the computing task R. These can be handled (in an analogous manner) according to the features of the method V, V1, V2.The at least one neural network DNN can be a plurality of neural networks DNN 10, DNN 11, DDD 12, DNN 13,... (which are possibly operated or operable in the backend BE and / or in one or more computing units 10, 11, 12, 13,... These may be interconnected or crosslinked with one another. The method is preferably carried out by means of a central neural network DNNZ. This can be pre-trained in the backend S20 and / or in the vehicles 10, 11, 12, 13,... and / or depending on vehicle data, re-trained S20'.In this case, a neural network DNN, DNNZ can be trained, in particular trained at night, to form a number N (e.g. more than 10, 100,... ) of sub-tasks R 1, R 2, R 3,... from a provided computing task R depending on the output layer of the neural network DNN, DNNZ; and / ormaking one or more sub-tasks R1, R2, R3,... executable on a plurality of computing units 10, 11, 12, 13,... ;executing S70 one or more sub-tasks R1, R2, R3 (exceptionally also whole computing tasks R) on a plurality of computing units 10, 11, 12, 13,... depending on the output layer of the neural network DNN, DNNZ, in particular to dynamically control S60.This (in relation to the one or more of steps S 40, S 50, S 60, S 70) is preferably carried out taking into account (e.g. for the best possible fulfilment) one or more target values (e.g. described in the present document), in particular target values related to a superordinate computing task R.For example, it can be decided (in a concrete case and / or in the case of a change) whether and which sub-tasks are formed from the computing task S 40, whether and how these are distributed S 50, whether and how they are executed S 70 controlled S 60. In other words, the method can (also) optimize the balance between the steps S 40, S 50, S 60, S 70, in particular whether and to what extent with which parameters the (respective) steps are to be carried out.For example, the one or more executions S 70 of the computing tasks R are operated S 30 (in particular also to be understood as controlled or influenced in their operation) depending on the operation S 30 of the one or more neural networks DNN(Z), DNN10, DNN11, DDN12, DNN13. These can be (interconnected) interconnected or mutually influencing neural networks DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13.For example, depending on output layers (of the neural networks DNN 10, DNN 11, DDN 12, DNN 13, DNN 13 assigned to the individual or specific computing units 10, 11, 12, 13, 14... or controlling them) a (further) forming S 40 of a number of sub-tasks from one or more sub-tasks R 1, R 2, R 3... depending on the output layer of a (respective) neural network DNN 10, DNN 11, DNN 12, DNN 13; and / ordistributing S50, controlling S60 and / or executing S70 one or more sub-tasks of the one or more of the sub-tasks R1, R2, R3,... are executed to a plurality of computing components, in particular belonging to the same computing unit 10, 11, 12 in each case, and / or belonging to one of the other computing units 13, 14, 15,...Operating S 30 the at least one neural network DNN, DNN 10, DNN 11, DNN 12, DNN 13, DNNZ can comprise or be controlling at least one step of the method by means of the respective at least one neural network DNN 1, DNN 2, DNNZ. For example, the at least one neural network (also to be understood as at least two interconnected neural networks DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13) can form part of the system S according to the invention.The method V can comprise or be (also only or with a combination of further features from the present document) training S 20 and / or operating S 30 the at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13) for carrying out one or more steps of the method.For different computing units 10, 11, 12, 13, 14, 15,... or selection of computing units 10, 11, 12, 13, respectively different, in particular specific training S20, in particular retraining S20' or different neural networks DNN, DNNZ, DNN10, DNN11, DNN12, DNN13) or operated with different learning data can be used.For example, the at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13) trained S 20, S 20' and / or operated S 30 can aim, in particular achieve, a target or target value relating to the computing unit 10, one or more further computing units 11, 12, in particular relating to a number of computing units 10, 11, 12, 13, 14, 15.For example, the at least one neural network DNN 10, DNNZ in the backend BE can be trained S 20, S 20' and / or operated S 30. Alternatively or additionally, the at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13) can be trained S 20, in particular retraining S 20' and / or operated S 30 in the respective computing unit 10, 11, 12, 13, 14, 15 and / or depending on the data of respective computing units 10, 11, 12, 13,...The at least one neural network DNN can be understood as data (e.g. topology, learning data, checkpoint, weight information, bias terms) for establishing the neural network DNN by means of at least one neural network DNN 1, DNN 2, DNNZ (e.g. on a data processing device). Therefore, a further neural DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13) in the sense of the present document is the same neural network DNN if it is or is at least predominantly depleted on a different computing unit.For example, the data are trained S20 on a data processing device in order to then make them retrainable S20' and / or operable S30 on one or more other data processing devices, in particular on (specific) computing units 10, 11, 12, 13,.., or backend BE and / or for the execution of computing tasks R on the computing units 10, 11, 12, 13, 14, 15,... and / or on specific computing units 10, 11, 12, 13.The at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13 may include or comprise one or more features of a convolutional neural network (CNN). The CNN is also to be understood in particular (according to corresponding German terms) as a convolutional neural network, a convolutional layer-comprising or so-called convolutional neural network. For example, the at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13 can comprise one or more convolutional layers (also to be understood as: convolutional layer).In training S 20, the core of the matrix may be learned. In this case, the core (for example with regard to the dimensioning) can be smaller than the input parameters. This core can be used for a plurality of cases, positions, various circumstances described herein, etc.The at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13 can recognize or learn influences, in particular structures of influences, that occur independently of case, location-independent and / or independently of one or more specific computing units and / or specific computing tasks.During operation S 30, these can be used, for example, for new cases or circumstances in each case. This can be carried out by means of a locally meshed layer (e.g. configured for this purpose), in particular the convolutional layer (convolutional layer).Furthermore, the information about the influences can be compressed in a matrix or transferred into a more abstract representation. In this case, at least predominantly in each case strongest features of a matrix can be prioritized or strengthened or the weaker features can be discarded. This can be done by means of the locally meshed layer (e.g. configured for this purpose), in particular pooling layer or subsampling layer. This can lead to a further reduction of the data to be further processed.Furthermore, the results from the two layers of the above-mentioned types mentioned above can be processed with one another, in particular can be combined analogously. In this case, the values received and / or the values received from the two slices can be processed together, in particular be calculated. This can be carried out in the third layer (or in the third step within the neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13). This can be an at least predominantly, in particular completely meshed layer of the at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13.Each of these optional features (e.g. of specific layers, the matrix of topology, etc.) of the one or more neural networks DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13 can be configured, dimensioned and / or (at least to a certain degree) optimized for the execution of at least one (corresponding) part of the method.The term "parameter" is also to be understood in the context of this document analogously as a parameter value or a measure of the change or deviation of a corresponding parameter value, for example of a change or deviation from a preceding value, a standard value or setpoint value. For example, a parameter can characterize a measure of the change in a parameter (to be understood as also: parameter value) over time and / or in relation to a (spatially variable) variable relating to the space, e.g. direction and / or distance. In particular, a parameter corresponds to a pattern (e.g. recognizable or executable as such) of parameters or parameter values.The term "characterizing" is also to be understood in particular as "representing" within the scope of the present document.The determination of one or more data mentioned in this document can in particular also be a prediction of the respective data, in particular with respect to one or more specific time intervals, e.g. of a few minutes, hours, days. Preferably (both) a determination (e.g. related to a value valid in the past or presence) and (also) a prediction (related to a future time interval, e.g. related to a specific time interval) are carried out.For example, the training S 20, S 20' and / or operation S 30 of the at least one neural network DNN, DNNZ, DNN 10, DNN 11, DNN 12, DNN 13 takes place as a function of data, in particular parameter (values) which are determined with respect to one or more (e.g. specific) time intervals and are predicted for one or more (e.g. specific) time intervals.FIGS. 3 and 4 illustrate (each) a variant of a plurality of mutually mutually mutually mutually mutually mutually combining variants of the invention.In one variant, the training S 20 of the at least one neural network DNN(Z) takes place in the backend BE.In this case, the training data can be determined on the basis of the first data and / or learning data of the at least one neural network DNN in the backend BE (the neural network is denoted there as DNN(Z). These can in turn be provided for training S 20, S 20' and / or operating S 30 or use in the one or more computing units 10, 11, 12, 13, 14, 15,... or for the computing units 10, 11, 12, 13, 14, 15,...For example, the learning data can be transmitted (directly or indirectly) to the respective computing unit 10 (e.g. in response to the sending of a multiplicity of data). The learning data (in particular to be understood as data of the at least one neural network DNN or for setting up the neural network DNN 1 can be loaded and / or loaded onto the one or more computing units 11, 12, 13, 14, 14 in order to operate the at least one neural network DNN therein depending on the learning data.In some variants of the invention which can be combined with all further features of the invention: the one or more of the computing units 10, 11, 12, 13, 14 can (in each case) be the computing units of one or more vehicles F.Said computing task R is understood here as a vehicle-external computing task R, in particular one provided from a (vehicle) backend BE, in particular one which is ordered.The one or more vehicles F are one or more at least partially, in particular fully electrically drivable vehicles F, for example a hybrid vehicle or a fully electrically drivable vehicle F.The one or more vehicles F may have (in each case) different states of charge of their (respective) batteries during the execution of the method V, V 1, V 2, e.g. may be at least partially charged and / or be in the state of charge. The information about this can be determined and taken into account in the method V, V 1, V 2.Alternatively or additionally, the state of charge and / or charging process (e.g. a parameter of the state of charge and / or of the charging process) and / or of the energy requirement can be predicted and taken into account for a time interval in the future. For example, if (e.g., for the next day) a high electrical energy requirement is predicted and, in particular, if the vehicle F is not connected to a charging station (also to be understood as: power grid, energy grid), the computing task R or sub-tasks R 1, R 2, R 3 can be initiated preferentially to other vehicles F or only one (tendingly small) sub-task R 1, R 2, R 3 can be initiated to the vehicle F.The vehicle F has hardware (within the scope of the document, in particular, to be understood as meaning or to be equivalent to computing units of different types and / or data memories). For example, they are one or more processors, in particular GPP and / or specialized or highly parallel processors, e.g. graphics processors.One or more parameters of the hardware installed in the one vehicle and / or resources of the hardware available for executing at least part of the computing task external to the vehicle are determined. In particular, these can be queried from the backend BE or read in there and taken into account.The vehicle 10 is configured for an operative connection (on both sides) with the backend BE for carrying out the method 100.The one or more computing units 10, 11, 12, 13 or (analogously) individual computing components of the computing units or computing components of the vehicles F can be computing units (hardware) of the (respective) vehicle F.The vehicle F can be designed to play software in a comparably flexible manner onto the one or more computing units or computing components of the vehicle 10, 11, 12, 13. The software for the operation of the respective vehicle F can be stored as a backup, while the one or more computing components are used to solve a computing task R. After the completion of the computing task R, the software for the operation of the vehicle F can then be installed again. The transmission of the software can take place via the communication unit of the vehicle F. In this case, the charging cable for charging an electrical energy store of the vehicle F and / or a wireless communication connection (e.g. 3G, 4G or 5G) can be used as the data connection.By means of a data exchange, a computing task R, in particular input data to a computing task R (also: for executing the computing task), can be read in, in particular called up, from outside the vehicle F. The computing task R can then be executed in the vehicle F, and the result E or E1, E2, E3 of the computing task R can be transmitted (e.g. to the same computing unit K by which the computing task R or R1, R2, R3 has been initiated or to a computing unit A networked thereto).The one or more vehicles F can be configured to perform an (information-related) operative connection with the vehicle-external computing unit (also: central unit or backend BE).In this case, an identifier and / or information for the available computing resources, for example nonvolatile memory, volatile memory, can be communicated (overall and / or in each case per control device) in the (respective) vehicle F. Alternatively or additionally, an identifier and / or information for the degree of use of the available resources for vehicle functions can be communicated. This can also be understood as information about the available resources. Alternatively or additionally, a forecast and / or time planning can be communicated with available resources of the (respective) vehicle.Furthermore, the method V, V 1, V 2 can comprise dividing a part of a computing task R 1, R 2, R 3 (also) within the vehicle F (transmitted to the respective vehicle F) into a plurality of computing units, in particular control units 10, 11, 12, 13 of the respective vehicle F; and / or into computing components (here: parts of the computing units, devices of the vehicle, control units) of the respective computing units 10, 11, 12, 13 (also to be understood as forming sub-tasks which each serve for processing the vehicle-external computing task or its parts).In particular, one or more read-in parts of the computing task R 1, R 2, R 3 can be divided among a plurality of computing units or computing components of the respective vehicle 10, 11, 12, 13. This can be done using the (communication) on-board power supply of the vehicle F. In this context, the computing resources available in each case in the individual computing units (e.g. control units) can be determined, for example by means of a query from outside the vehicle and / or inside the vehicle F.The division of a computing task R into the different vehicles F and / or computing components 10, 11, 12, 13 of the (respective) vehicle can then be carried out in accordance with the type of required computing operations (e.g. GPU= Graph Processing Unit), parallel processing, graphics processing, neural processing).One or more of the plurality of computing units 10, 11, 12, 13,... can be configured with a first computing mode which (for example in contrast to an operating mode which is configured in particular for said first computing mode and is configured for executing the functionality of the respective device), supports improved feasibility of the external computing tasks and / or the execution of the method according to one or more features.Putting the computing unit 10, 11, 12, 13 into the first computing mode can comprise flashing at least partially new firmware onto one or more computing units (in the case of a vehicle, for example, onto control units of the respective vehicle F). In this case, (directly or indirectly) the resources of one or more EEPROMs (electrically erasable programmable read-only memory) of the computing unit (in an improved manner for the method V, V 1, V 2) can also be used.The one or more computing units 10, 11, 12, 13,... can be used, in particular, for an executable program code which differs from the functional software of the respective devices. The program code can be part of the computing task R (or a sub-task R 1, R 2, R 3). Input data relating to the computing task R (or a sub-task R 1, R 2, R 3) can be stored on the one or more computing units. Alternatively or additionally, the one or more computing units 10, 11, 12, 13,... can be used as a medium-term or long-term buffer or for memory management for the (global) computing task R. One or more vehicles F can thus be used to store (such as vehicle-external and / or vehicle-independent) data.For example, one or more vehicles F may be configured to execute a second computing mode usable during vehicle operation, in particular during driving operation. The second calculation mode can be enabled, for example, during driving operation on specific roads, locations, countries, etc. The second calculation mode can be provided depending on the degree of use and / or on the prediction of use of vehicle functions of the (respective) vehicle F. It is thus possible to allow at least a portion of the computing resources of a vehicle F to be used during vehicle operation.Further features are described below which can be freely combined with one another and / or with all further features described in the present document.The method V, V 1, V 2 may comprise a "brokerage" and / or "contracting" with respect to a computing task and / or sub-tasks R 1, R 2, R 3. In particular, the method can comprise comparing one or more offers from one or more computing units 10, 11, 12, 13 with respect to a reward.For example, a comparison or balancing, in particular an optimization, of the acceptance of a computing task R or sub-task R 1, R 2, R 3 can take place. For example, it can be weighed whether a larger computing task R is accepted for a longer time for less counter-power (e.g. money units, counter-power units, triggerable conditions) per computing power or a smaller computing task or sub-task R1, R2, R3 with more counter-power or money per computing power.For example, the at least one neural network may be DNN, DNN 10, DNN 11, DD 12, DN 13,... DNNZ can be configured, in particular trained, and / or operated to perform the brokerage and / or contracting with respect to a computing task and / or sub-tasks R 1, R 2, R 3 (e.g., among each other). In this case, it can optimize it to a target value (for example, changeable or predefinable by a user), in particular a result of weighting a plurality of target values.Alternatively or additionally, the one or more computing units 10, 11, 12, 13 can be configured to send and / or adapt a separate offer for accepting a computing task R or a sub-task R 1, R 2, R 3 (e.g. to the central unit BE). Furthermore, the backend BE and / or the one or more vehicles 10, 11, 12, 13 can be configured to bring about an automatic completion of an agreement with respect to the execution of a computing task R or of one or more sub-tasks R 1, R 2, R 3.The invention or measures described in this document can substantially increase the utilization of the computing resources of a computing unit 10, 11, 12, 13.Thus, the effective costs of a computing unit 10, 11, 12, 13,... in particular costs for operating the computing unit or total costs, in particular by marketing and / or using the computing resources for external computing tasks, e.g. vehicle-independent computing tasks), can be reducedThe present invention is not limited to the exemplary embodiments shown. In particular, it should be noted that the description and the figures are intended to illustrate, by way of example only, the principle of the proposed methods, apparatuses and systems. All features identified in the exemplary embodiments shown can be freely combined with all other features or claims.

Claims

Method (V, V1) for executing (S70) a computing task (R) on a plurality of computing units (10, 11, 12, 13), comprising - forming (S40) a number of sub-tasks (R1, R2, R3,... ) from a provided computing task (R) as a function of the output layer of a neural network (DNN, GPT); and - distributing (S50), from the formed sub-tasks (R1, R2, R3,... ) to a plurality of computing units (10, 11, 12) for achieving (S70) a result (E, E1, E2, E3,... ) of the computing task (R).Method (V, V2) for executing (S70) a computing task (R) on one or more computing units (10, 11, 12, 13), comprising - training (S20), in particular retraining (S20'), a neural network (DNN, GPT) from a computing task (R) to form (S40) a plurality of sub-tasks (R1, R2, R3), in particular dividing the computing task (R) into a plurality of sub-tasks (R1, R2, R3), in such a way that one or more sub-tasks (R1, R2, R3,...) can be executed (S70) on a plurality of computing units (10, 11, 12, 13).Method (V, V1) according to Claim 1, wherein distributing (S40) the one or more sub-tasks (R1, R2, R3,...) to a plurality of computing units (10, 11, 12, 13) comprises or is controlling (S60) the execution (S70), in particular dynamic and / or adaptive controlling (S60) the execution (S70) of the one or more respective sub-tasks (R1, R2, R3,...).Method (V, V2) according to Claim 3, wherein the controlling (S60) of the execution (S70) of the one or more sub-tasks (R1, R2, R3,...) is or comprises: - controlling (S60) an interface provided in the one or more sub-tasks (R1, R2, R3,...), in particular by means of the interfaces provided in the executable code of the respective sub-tasks (R1, R2, R3,...); and / or - controlling (S60) the operating system of the respective computing units (10, 11, 12, 13), wherein the operating system influences the execution (S70) of the respective sub-tasks (R1, R2, R3,....) as a function of the controlling (S40).Method (V, V1, V2) according to one of the preceding claims, wherein - the distribution (S50) of one or more sub-tasks (R1, R2, R3,...) on a plurality of computing units (10, 11, 12, 13) is or comprises redistribution (S50') of one or more of the sub-tasks (R1, R2, R3,...), in particular during the handling of the computing task (R) and / or during the execution of one or more of the sub-tasks (R1, R2, R3,...).Method (V, V1, V2) according to one of the preceding claims, wherein the control (S60) of the execution (S70) of a first sub-task (R1, R2, R3,...) is carried out on a first arithmetic unit (10, 11, 12, 13,...): - as a function of the control (S60) of the execution (S70) of one or more second sub-tasks (R1, R2, R3,...) on one or more second arithmetic units (10, 11, 12, 13); and / or - as a function of the progress, intermediate result and / or result of the execution (S70) of one or more second sub-tasks (R1, R2, R3,....) on one or more second arithmetic units (10, 11, 12, 13).Method (V, V1, V2) according to one of the preceding claims, wherein the at least one neural network (DNN, GPT) comprises: - a generative pre-trained transformer (GPT); and / or - a generative, in particular statistical model; and / or - a statistical model trained at specific computing units, users of the devices comprising the computing units (10, 11, 12, 13), computing tasks; can be constructed on the basis of such a model and / or is controlled by such a model.Method (V, V1, V2) according to one of the preceding claims, wherein the at least one neural network (DNN, GPT) is trained (S20), in particular trained at night (S20'), to aim at, in particular at least presumably to reach and / or to hold, a target value with respect to the calculation task (R).Method (V, V1, V2) according to one of the preceding claims, wherein the training (S20), in particular night training (S20'), and / or operation (S30) of the at least one neural network (DNN, GPT) is carried out as a function of one or more parameters with respect to the one or more arithmetic units (10, 11, 12, 13, 14, 15,..).Method (V, V1, V2) according to one of the preceding claims, wherein the at least one neural network (DNN, GPT) is or is trained (S20), in particular retraining (S20'), in particular retraining (S20'), depending on the computing power available and / or used in the past of one or more of the plurality of computing units (10, 11, 12, 13, 14, 15,...) or of a higher-order number of computing units (10, 11, 12, 13, 14, 15,..).Method (V, V1, V2) according to Claim 6, wherein the at least one target value relates to a cost required for executing (S70) at least part of the computing task (R), in particular a number of sub-tasks (R1, R2, R3), in particular: - power costs for executing (S70) at least part of the computing task (R), in particular all essential parts of the computing task (R); and / or - data transmission costs for transmitting at least part of the computing task (R), in particular a number of the plurality of sub-tasks (R1, R2, R3), to one or more computing units (10, 11, 12, 13); and / or data transmission costs for transmitting at least a part of the one or more calculation results (E1, E2, E3), in particular all essential calculation results (E1, E2, E3), from one or more calculation units (10, 11, 12, 13), in particular within the scope of providing the overall result of the calculation task (R).Method (V, V1, V2) according to one of the preceding claims, comprising - making, in particular negotiated, an agreement relating to a compensation and / or a counter-performance for the execution of the computing task (R) on one or more of the plurality of computing units (10, 11, 12, 13,...) or the devices (F) comprising the computing units (R); and - forming (S40) a number of sub-tasks (R1, R2, R3,...) from a computing task (R); and / or - distributing (S50) sub-tasks (R1, R2, R3,...) from the number of sub-tasks (R1, R2, R3,...) to a plurality of computing units (10, 11, 12, 13); and / or - controlling (S 60) the execution (S 70) of a first sub-task (R 1, R 2, R 3,... ) on the one or more computing units (10, 11, 12, 13); depending on the agreement made.Method (V, V1, V2) according to Claim 12, comprising - carrying out one or more steps of the method (V, V1, V2) according to Claim 12 by means of a neural network (DNN, GPT), in particular of the neural network (DNN, GPT) or a bot (DNN, GPT) operated by the neural network, wherein the neural network (DNN, GPT) - training (S20) and / or operating (S30) the neural network (DNN, GPT) to minimize the compensation and / or the counter power for the execution of the computing task (R) on one or more of the plurality of computing units (10, 11, 12, 13,...); wherein the training (S 20) is or has been carried out in particular as a function of individual compensations and / or counter-powers for computing units (10, 11, 12, 13,... ) and / or the entire compensation and / or counter-power for one or more computing tasks (R) in the past; and / or the operating (S 30) is or has been carried out in particular as a function of individual compensations and / or counter-powers for computing units (10, 11, 12, 13,... ).Method (V, V1, V2) according to one of the preceding claims, comprising, - executing (S70) one or more sub-tasks (R1, R2, R3,...) on one or more computing units (10, 11, 12, 13,...) by means of a further computer program which is used for operating and / or optimizing one or more computing components of the one or more computing units (10, 11, 12, 13) to execute (S70) the one or respective one of the plurality of sub-tasks (R1, R2, R3), wherein the computer program, in particular in the form of firmware, is applied to and / or operated on the one or in each case a plurality of computing units (10, 11, 12, 13,..).Method (V, V1, V2) according to one of the preceding claims, comprising: - receiving a plurality of results (E1, E2, E3) from the plurality of sub-tasks (R1, R2, R3), and determining a total result (E) for the computing task (R) depending on the plurality of results (E1, E2, E3); and - providing the total result (E) to the computing task (R).Method (V, V1, V2) according to one of the preceding claims, wherein - the at least one computing task (R) is or comprises a computing task from a vehicle (F), and the one or more computing units (10, 11, 12, 13) are computing units (10, 11, 12, 13) outside the vehicle (F); or - at least one computing task (R) is or comprises a vehicle-external computing task (R), and the one or more computing units (10, 11, 12, 13) are computing units (10, 11, 12, 13) of one or more vehicles (F1, F2, F3); or - the plurality of computing units (10, 11, 12, 13) are different computing units (10, 11, 12, 13) of the same vehicle (F).Neural network (DNN, GPT), wherein the neural network (DNN, GPT) is configured, in particular trained (S20) and / or retraining (S20') or retraining (S20') to form (S40) a plurality of sub-tasks (R1, R2, R3) from a computing task (R), in particular to divide the computing task (R) into a plurality of sub-tasks (R1, R2, R3) such that one or more sub-tasks (R1, R2, R3,..) can be executed (S70) on a plurality of computing units (10, 11, 12, 13).System (S) for executing (S70) a computing task (R) on one or more computing units (10, 11, 12, 13), wherein the system (S) is configured to - form (S40) a number of sub-tasks (R1, R2, R3,... ) from a computing task (R) depending on the output layer of a neural network (DNN, GPT); and is configured to: - distribute (S50) the formed sub-tasks (R1, R2, R3,... ) to a plurality of computing units (10, 11, 12) for achieving (S70) a result (E, E1, E2, E3,... ) of the computing task (R).System (S) for executing (S70) a computing task (R) on one or more computing units (10, 11, 12, 13), wherein the system (S) is configured to form (S40) a plurality of sub-tasks (R1, R2, R3) from a computing task (R) for training (S20), in particular retraining (S20'), a neural network (DNN, GPT) in such a way that one or more sub-tasks (R1, R2, R3) can be executed (S70) on a plurality of computing units (10, 11, 12, 13) for achieving (S70) a result (E, E1, E2, E3,...) of the computing task (R).Computer program, in particular a computer program product, comprising a computer program, wherein the computer program is designed, when executed on at least one arithmetic unit, in particular on one or more arithmetic units of the system (S) according to Claim 17 or 18, to execute one or more steps of the methods (V, V1, V2) according to at least one of Claims 1 to 16.

Images