Global interconnection Isin architecture computing system based on DRAM (Dynamic Random Access Memory)
By introducing Ising control module and forward addition module into the DRAM storage controller, the globally interconnected Ising architecture computing system is realized, solving the problems of insufficient spins and high power consumption in the existing technology, and realizing Efficient combination optimization solution and low-cost global interconnection.
Patent Information
- Application Number
- CN202410861074.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-28
- Publication Date
- 2024-11-08
Smart Images

Figure CN118917426A_ABST
Abstract
Description
Technical Field
[0001] The present invention mainly relates to the field of computer technology, and in particular to a DRAM-based global interconnected Ising architecture computing system. Background Art
[0002] The current Ising architecture computing system has a small number of spins, especially the number of spins of the Ising architecture computing system that realizes global interconnection. The Ising architecture computing system that realizes global interconnection has a small number of spins, which makes it difficult to solve large-scale combinatorial optimization problems. However, the mapping and solution of most combinatorial optimization problems require a globally connected Ising architecture. Although current researchers have made innovations in structures such as spin updaters and reduced some overhead, they still mainly expand the CMOS Ising architecture computing system by increasing the storage capacity of SRAM.
[0003] Since most Ising chips rely on SRAM storage, the increase in spin scale will inevitably increase the capacity of SRAM, and the increase in SRAM capacity will bring non-negligible power consumption and area overhead. Therefore, it is still difficult for the current Ising chip to realize large-scale spin systems. The capacity of the spin system built based on SRAM is limited, and the scale of the problems that can be solved is often relatively small, which limits its practicality.
[0004] The latest research progress has achieved the Ising architecture computing system based on eDRAM, but it only realizes the local interconnection of adjacent 8 spins, and still has not realized the global interconnection system, so the types of problems that can be solved are relatively small. Summary of the invention
[0005] The technical problem to be solved by the present invention is: in view of the deficiencies in the prior art, a DRAM-based globally interconnected Ising architecture computing system is provided which has low cost, high interconnectivity and high solving speed.
[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions: A DRAM-based global interconnected Ising architecture computing system comprises a DRAM storage controller, wherein an Ising control module is provided in the DRAM storage controller, wherein the Ising control module comprises an annealing updater, a random source and an approximate probability flip term generator, so as to realize the control of annealing calculation by the storage controller; the DRAM storage controller adopts a carry-lookahead addition module inside a storage array to perform update calculation of local search terms, and designs an SHF16 row to realize synchronization of local search terms.
[0007] As a further improvement of the present invention, in the Ising control module, each spin node is connected to each other in pairs; each spin node is used to store the interaction coefficient with all other spins, and when the interaction coefficient bit width is set to 16 bits, the bit width of the local search item is set to 32 bits to prevent data overflow.
[0008] As a further improvement of the present invention, each sub-array partial storage row in the DRAM storage controller is divided into five parts as a functional row: (1) Group C is a constant group; C0 and C1 store 0 and 1 respectively, forming constant 0 and constant 1 rows, and C2 stores a 32-bit wide "+1" complement, i.e. "32'h00000001"; (2) Addend row A; A is another addend, which is the product of the interaction coefficient between the spin and the selected spin and the new spin state of the selected spin; (3) Local search item LC row; LC is the value of the local search item of the spin; A and LC together constitute the addend part of the adder; (4) SHF16 line; SHF16 is used to synchronize local search items; (5) Addition module function lines, intermediate calculation reserve lines and function lines required when the addition operation is running.
[0009] As a further improvement of the present invention, a part of Ising control logic is provided in the DRAM storage controller to realize the annealing process and spin update calculation of the control system; the Ising control module includes three modules: a random number generator RandomNum, a spin updater Updater and a probability flip item generator ProbFlip.
[0010] As a further improvement of the present invention, the spin node is composed of a "storage+computation" subarray or a "storage+computation" subarray and multiple "storage" subarrays.
[0011] As a further improvement of the present invention, if the spin state of the selected spin node changes, the local search term is directly added or subtracted by the double interaction coefficient corresponding to the selected spin. To update, it is as follows: .
[0012] As a further improvement of the present invention, the storage mapping mode is to store only Interaction coefficient.
[0013] As a further improvement of the present invention, the storage mapping method is: storing and ; Since the spin value has only two possible values, +1 and -1, we can calculate and , based on the positive or negative selection of the latest state of the selected spin or , and then perform addition calculation to obtain a new local search item.
[0014] As a further improvement of the present invention, each spin node is numbered in sequence with numbers, and the interaction coefficients between the spin and each spin are stored in the storage subarray of each spin node in order from left to right and from top to bottom according to the size of the spin node number, and "0" is stored in the storage position corresponding to the spin node's own number.
[0015] As a further improvement of the present invention, reading the interaction coefficient of the selected spin is divided into two cases. One is that the interaction coefficient of the selected spin is located in the "storage + calculation" subarray, and the "AAP" operation is used to copy it to the addend row to complete the reading operation; the other case is that the interaction coefficient of the selected spin is located in the "storage" subarray, and copying it to the "storage + calculation" subarray is a cross-array operation. By adding isolation transistors, connecting adjacent subarrays, and utilizing the ability to quickly move data between subarrays in DRAM, the interaction coefficient of the specified spin can be read.
[0016] Compared with the prior art, the advantages of the present invention are: 1. The Ising architecture computing system based on DRAM to achieve global interconnection of the present invention has low cost, high interconnection degree and high solving speed. The present invention provides two storage mapping modes. The normal mode only stores positive two-fold interaction coefficients, and the pre-calculation mode stores positive and negative two-fold interaction coefficients respectively.
[0017] 2. The Ising architecture computing system of the present invention realizes global interconnection based on DRAM, wherein each spin node is composed of one or more sub-arrays. By improving the storage controller, it is possible to effectively control the completion of annealing calculations, thereby completing the solution to the combinatorial optimization problem. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is the overall structural principle diagram of the global interconnection system based on DRAM of the present invention.
[0019] Figure 2 Schematic diagram of specific row composition of the “storage+computation” subarray in a specific embodiment of the present invention.
[0020] Figure 3 It is a schematic diagram of the abstract structure of the global interconnection system in a specific embodiment of the present invention.
[0021] Figure 4Schematic diagram of the annealing step of the De Gloria method in a specific embodiment of the present invention.
[0022] Figure 5 Schematic diagram of a subarray of pre-calculated patterns in a specific embodiment of the present invention.
[0023] Figure 6 It is a schematic diagram of the process of solving the combinatorial optimization problem by the system in a specific embodiment of the present invention.
[0024] Figure 7 It is a schematic diagram of a specific process of updating respective local search items in parallel in a specific embodiment of the present invention.
[0025] Figure 8 It is a schematic diagram of the mapping relationship between addresses and word lines in a specific embodiment of the present invention. DETAILED DESCRIPTION
[0026] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0027] The present invention provides a DRAM-based global interconnected Ising architecture computing system. The present invention adds an Ising control module to a storage controller, the module includes three components: an annealing updater, a random source, an approximate probability flip term generator, and the like, so as to control the annealing calculation using the storage controller.
[0028] In a specific application example, the present invention uses a carry-lookahead addition module inside a storage array to perform update calculations of local search items, and designs a SHF16 row to achieve synchronization of local search items.
[0029] The globally interconnected Ising architecture computing system implemented based on DRAM in the present invention has a large spin scale, low cost, and high speed, is suitable for solving larger-scale and more complex combinatorial optimization problems, and is also more conducive to the application and promotion of the Ising architecture computing system.
[0030] like Figure 1-Figure 3 As shown, the overall structure of the global interconnected Ising architecture computing system based on DRAM is demonstrated. The main innovation of the present invention is that an Ising control module is added to the DRAM storage controller, and the Ising control module is used to implement spin update control and random flip item generation. The bottom few rows of each sub-array are the logic rows required for spin update calculation.
[0031] In the Ising architecture that implements global interconnection, each spin is connected to each other, which requires each spin node to store the interaction coefficients with all other spins. When the interaction coefficient bit width is set to 16 bits, in order to avoid data overflow, the bit width of the local search item is set to 32 bits.
[0032] like Figure 2 As shown, in a specific application example, for a subarray of size 512*512, a row of 512 storage cells can store 16 local search items. Two addends are stored horizontally in the storage cells. When performing addition calculations, the lengths of the two addends must be equal, so each storage row can hold the interaction coefficients corresponding to 16 spins.
[0033] Each subarray part stores a row as a function which can be divided into 5 parts: (1) Group C is a constant group. C0 and C1 store 0 and 1 respectively, forming a constant 0 and constant 1 row. C2 stores a 32-bit wide "+1" complement, i.e., "32'h00000001".
[0034] (2) Addend row A; A is another addend, that is, the product of the interaction coefficient between the spin and the selected spin and the new spin state of the selected spin.
[0035] (3) Local search term LC row; LC is the value of the local search term for this spin. A and LC together constitute the addend part of the adder.
[0036] (4) SHF16 line; SHF16 is used to synchronize local search items.
[0037] (5) Addition module function lines. The intermediate calculation lines and function lines required when the addition operation is running are reserved.
[0038] The subarrays are divided into "storage + computing" subarrays and "storage" subarrays. Figure 2 As shown, it contains the computational function rows, and the "storage" subarray is used only to store the interaction coefficients with other spins.
[0039] In a specific application example, the present invention adds a part of Ising control logic to the storage controller to realize the annealing process and spin update calculation of the control system. The Ising control module mainly includes three modules: a random number generator RandomNum, a spin updater Updater and a probability flip item generator ProbFlip. The random number generator is used to generate random numbers for subsequent random selection of spin updates. The spin updater is used to control the annealing process of solving the problem. The probability flip item generator is used to realize the generation of probability flip items in the spin update calculation. Depending on the scale of the problem, the spin node is composed of a "storage + calculation" subarray or a "storage + calculation" subarray and multiple "storage" subarrays.
[0040] In order to obtain the next spin state of the spin, the present invention first needs to calculate the local search item, and then, it needs to flip the spin state probabilistically so that the problem may jump out of the local optimal solution. The calculation of the local search item is completed inside the spin node, and each spin node can update its own local search item in parallel, see Figure 7 .
[0041] As a preferred embodiment, see Figure 4 ,The present invention adopts the De Gloria method.,Each time the spin state is updated iteratively, the system does not need to re-evaluate the overall interaction between all spins, but can only rely on the local search items of each spin to perform local optimization search.
[0042] In order to improve efficiency, the present invention further provides a "storage + calculation" subarray in each spin node of the designed system, and the update of local search items is completed in the "storage + calculation" subarray. When the state of the selected spin changes, the local search items of other spins can be updated independently in their "storage + calculation" subarrays in parallel. This design greatly optimizes the local search time of the spin node.
[0043] Probability flipping is to ensure that in the process of updating the spin state, the system will not always stay in a local minimum energy state and cannot be further optimized. The present invention further adopts Gibbs sampling in the probability flipping model of the Ising model. The probability flipping term is randomly generated, which allows a certain degree of error within an acceptable range of accuracy. The approximate probability flipping calculation technology cleverly utilizes the above characteristics to achieve the generation of probability flipping terms with low hardware overhead.
[0044] As a preferred embodiment, the present invention uses an approximate probability flip calculation technique to realize the generation of probability flip terms.
[0045] In order to save the calculation time of local search items and facilitate the control of spin annealing calculation, on the one hand, the calculation of local search items needs to be simplified as much as possible. On the other hand, the spins need to be stored in a cell array in a regular manner.
[0046] According to the local search term formula, since the spin state has only two cases, spin up or spin down, that is, the spin value is "+1" and "-1". Therefore, if the spin state of the selected spin changes, the local search term can be directly added or subtracted by the double interaction coefficient corresponding to the selected spin. To update, as follows:
[0047] Therefore, in order to facilitate calculation, the present invention directly stores the double interaction coefficient in the storage unit .
[0048] In a specific application process, according to the size of the system, the present invention provides two storage mapping modes.
[0049] The first is to store only Interaction coefficient.
[0050] The other is to store separately and Since the spin value has only two possible values, +1 and -1, we can calculate and , based on the positive or negative selection of the latest state of the selected spin or Then, the new local search item can be obtained by performing addition calculation. In the system of the present invention, only the positive double interaction coefficient is stored, which is called the normal mode, and the mode of storing the positive and negative double interaction coefficients separately is called the pre-calculation mode.
[0051] See also Figure 5 For the pre-calculation mode, the specific storage method is to store even rows. , odd-numbered rows are stored , the calculation function area is the same as the normal storage mode.
[0052] It should be noted that the pre-calculation mode is recommended for solving relatively small-scale combinatorial optimization problems, that is, the size of a spin node containing only one "storage + calculation" subarray. Otherwise, since the required storage space is twice that of the normal mode, this method sacrifices a certain amount of storage space.
[0053] If it is stored in normal mode, only the positive double interaction coefficient is stored. , so if the new state of the selected spin is "+1", just add twice the interaction coefficient; if the new state of the selected spin is "-1", since the interaction coefficient is stored in the form of complement code in this system, it is only necessary to complete the calculation according to the following formula:
[0054] That is Take the complement of the inverted code and add one to get The complement of , and then add the original local search item to get the latest local search item.
[0055] From the above, we can see that Figure 6The system of the present invention solves the combinatorial optimization problem by first mapping the problem to the Ising model, then storing the model coefficients and local search items in the storage unit, and then configuring the spin update calculation control information to start solving the combinatorial optimization problem. The Ising annealing process controller is improved based on the original DRAM storage controller, and mainly adds three parts, namely, the random number generator, the probability flip term generator and the updater.
[0056] The spin update calculation is to add the local search term to the probability flip term. The probability flip term is randomly generated, so the local search term allows a certain error. Secondly, the design of the storage controller prefetch mechanism reads the entire target row when reading data. In addition, in order to simplify the calculation, the interaction coefficient Multiply by 2 and store again, The lowest bit of the local search item is always "0", so the lowest bit of the original local search item will not be changed after the local search item is updated. In summary, the present invention writes the spin state in the lowest bit of the local search item, and when reading the local search item LC row, it is equivalent to reading its local search item and spin state. If the state of the selected spin is updated, it is necessary to change the lowest bit of the local search item, and then write the new local search item back to the spin node local search item storage row.
[0057] The controller randomly selects a spin a from all the spins, and reads the local search item LC row of spin node a to the storage controller. Next, the local search item is added to the probability flip item, and the spin value is determined based on the positive or negative result after the addition result is obtained. The spin state is written to the lowest bit of the local search item to obtain a new local search item, and it is written back to the local search item storage row of the spin node. If the spin state changes, the other spin nodes are notified to update their respective local search items, and the step number STEP is set to 0, otherwise the step number STEP+1. If the step number is less than the set step number, a spin is selected by the random source again, and the above steps are repeated until the step number is greater than the set step number, then the annealing can be considered complete, and the spin state, that is, the lowest bit of the local search item of each spin node, is obtained.
[0058] If the selected spin a is updated, the local search items of other spins will change, so the local search items of other spins need to be recalculated according to the formula: (1) First, the storage controller reads the corresponding interaction coefficient according to the number of the selected spin node. Figure 2 Then, depending on the latest state of the selected spin, decide whether to perform one addition or two additions. If the latest state of the selected spin is 0, then Adding the local search item will give a new local search item; if the latest state of the selected spin is 1, then The new local search item is obtained by inverting and adding 1, and then adding the local search item. The addition is calculated by the addition operation module, and finally the local search item is synchronized through the SHF16 line.
[0059] All spins in the global interconnected system are interconnected. If the selected spin has a state update, other spins need to update their local search items. When updating a local search item, the interaction coefficient corresponding to the updated spin needs to be read first, then added to the original local search item to obtain a new local search item, and finally the local search item is updated synchronously.
[0060] In order to facilitate memory-controlled spin annealing, in a preferred embodiment, the system of the present invention numbers each spin node in sequence with a number, and stores the interaction coefficients of the spin with each spin in the storage subarray of each spin node in order from left to right and from top to bottom according to the spin node number size, and stores "0" in the storage position corresponding to the spin node's own number. Figure 2 and Figure 7 , for spin node 0, then at row 0, column 0 The position stores "0", at row 0, column 1 The interaction coefficient between the position storage and the spin node 1. According to the method described above, the interaction coefficient between each spin node and a certain spin node is in the same relative position, so the storage controller can also easily control the interaction coefficient between each spin node and the selected spin node.
[0061] In a globally interconnected system, if the selected spin changes its spin state, all other spins need to recalculate their local search terms. The local search term calculation needs to be calculated according to equation (1). The value of has only two cases, "+1" and "-1", so the calculation of the local search term can be simplified to the negation operation and the addition operation.
[0062] The inversion and addition calculations of the globally interconnected CMOS Ising architecture computing system implemented by the present invention are mainly based on an improved carry-lookahead adder based on DRAM, so that it can be applied to modern high-density DRAM design. The addition calculation requires AND, OR, XOR and other operations, which require fine-tuning of the decoder. Figure 8, through the mapping relationship between addresses and word lines, such as address B0 represents mapping R0 row, and address R8 simultaneously maps R0 and R3. For example, AP(B0) represents activating R0 row, and AP(R8) represents activating R0 and R3 rows at the same time. The first 6 addresses activate the word lines of 6 rows R0-R6 respectively, B7 activates the word line of NOT row, B8, B9, B10 activates 2 word lines at the same time, B11 activates 4 word lines at the same time, B12 to B16 and B18 activate 3 word lines at the same time, B17 activates the word line of Shift row to implement the operation of transferring carry, B19 activates PC and Shift word lines to generate carry and save the carry to Shift row. A address activates the storage row storing another addend, and LC activates the storage row storing local search terms. C0, C1, and C2 addresses can activate three storage rows storing fixed numbers respectively.
[0063] Further, as a preferred solution, the present invention reads the interaction coefficient of the selected spin in two cases. One is that the interaction coefficient of the selected spin is located in the "storage + calculation" subarray. In this case, the "AAP" operation is used to copy it to the addend row to complete the reading operation. The other case is that the interaction coefficient of the selected spin is located in the "storage" subarray. Copying it to the "storage + calculation" subarray is a cross-array operation. By adding isolation transistors and connecting adjacent subarrays, the ability to quickly move data between subarrays in DRAM can be used to quickly read the interaction coefficient of the specified spin. Depending on the size of the system, the delay across multiple subarrays is different from the delay across one subarray.
[0064] When calculating the local search term, there are two cases that need to be considered. One case is that the latest state of the selected spin is spin-up, that is, the spin value is "+1". In this case, only one addition operation is required to calculate That is, it is called "positive state addition". Among them, AAP(X,Y) means that X is first activated for copy or multi-row activation operation, and then the sense amplifier is enabled until a stable result is obtained. Then, the word line of X is turned off, and Y is activated to perform a write-back operation. AP(X) means activating X, and then turning off the word line of X after enabling the sense amplifier to obtain a stable result.
[0065] Another case is that the latest state of the selected spin is spin-down, that is, the spin value is "-1". In this case, two addition operations need to be completed, which is called "negative state addition". The first is to add After negation, add 1, and then add L(i) to calculate. Although the specific operation instructions of these two cases are similar, there are differences in the addend part.
[0066] For a 512*512 size subarray, the storage capacity of each row is sufficient to store 16 32-bit wide local search items. During a spin update, only one of the local search items is effectively calculated. Therefore, after the calculation is completed, the new local search item needs to be synchronized to the other local search items in the row. Figure 2 As shown, the SHF16 row designed by the present invention is to add a row of transistors on the basis of the DRAM storage row. This design solves the synchronization problem and ensures that all local search items can be updated promptly and accurately after each spin update.
[0067] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. A DRAM-based global interconnected Ising architecture computing system, comprising a DRAM storage controller, It is characterized in that The DRAM storage controller is provided with an Ising control module, which includes an annealing updater, a random source and an approximate probability flip term generator, so as to realize the control of annealing calculation by the storage controller; the DRAM storage controller uses a carry-lookahead addition module inside the storage array to perform update calculation of local search terms, and realizes synchronization of local search terms through SHF16 rows.
2. The DRAM-based global interconnected Ising architecture computing system according to claim 1, Features: In the Ising control module, each spin node is connected to each other in pairs; each spin node is used to store the interaction coefficient with all other spins. When the interaction coefficient bit width is set to 16 bits, the bit width of the local search item is set to 32 bits to prevent data overflow.
3. The DRAM-based global interconnected Ising architecture computing system according to claim 2, Features: Each sub-array portion storage row in the DRAM memory controller is divided into five parts as a functional row: (1) Group C is a constant group; C0 and C1 store 0 and 1 respectively, forming constant 0 and constant 1 rows, and C2 stores a 32-bit wide "+1" complement, i.e. "32'h00000001"; (2) Addend row A; A is another addend, which is the product of the interaction coefficient between the spin and the selected spin and the new spin state of the selected spin; (3) Local search item LC row; LC is the value of the local search item of the spin; A and LC together constitute the addend part of the adder; (4) SHF16 line; SHF16 is used to synchronize local search items; (5) Addition module function lines, intermediate calculation reserve lines and function lines required when the addition operation is running.
4. The DRAM-based global interconnected Ising architecture computing system according to claim 2 or 3, Features: The DRAM storage controller is provided with a part of Ising control logic for realizing the annealing process and spin update calculation of the control system; the Ising control module includes three modules: a random number generator RandomNum, a spin updater Updater and a probability flip item generator ProbFlip.
5. The DRAM-based global interconnected Ising architecture computing system according to claim 4, Features: The spin node is composed of a "storage+computation" subarray or a "storage+computation" subarray and multiple "storage" subarrays.
6. The DRAM-based global interconnected Ising architecture computing system according to claim 5, Features: If the spin state of the selected spin node changes, the local search term is directly added or subtracted by the double interaction coefficient corresponding to the selected spin. To update, it is as follows: 。 7. The DRAM-based global interconnected Ising architecture computing system according to claim 6, Features: The storage mapping mode is to store only Interaction coefficient.
8. The DRAM-based global interconnected Ising architecture computing system according to claim 6, Features: The storage mapping method is: store separately and ; Since the spin value has only two possible values, +1 and -1, we can calculate and , based on the positive or negative selection of the latest state of the selected spin or , and then perform addition calculation to obtain a new local search item.
9. The DRAM-based global interconnected Ising architecture computing system according to claim 6, Features: Each spin node is numbered in sequence, and the interaction coefficients between the spin and each spin are stored in the storage subarray of each spin node in order from left to right and from top to bottom according to the spin node number size, and "0" is stored in the storage position corresponding to the spin node's own number.
10. The DRAM-based global interconnected Ising architecture computing system according to claim 6, Features: There are two cases for reading the interaction coefficient of the selected spin. One is that the interaction coefficient of the selected spin is located in the "storage + computation" subarray, and the "AAP" operation is used to copy it to the addend row to complete the reading operation. The other case is that the interaction coefficient of the selected spin is located in the "storage" subarray, and copying it to the "storage + computation" subarray is a cross-array operation. By adding isolation transistors, connecting adjacent subarrays, and utilizing the ability to quickly move data between subarrays in DRAM, the interaction coefficient of the specified spin can be read.
Citation Information
Cited By
Isin machine circuit based on ALCU model and adaptive AMV
CN119724280A
