Method and apparatus for processing input quantities
The use of FETs and FeFETs to selectively connect transistor load sections with circuit nodes addresses inefficiencies in processing binary and ternary data, enabling efficient current signal conversion and acceleration of neural network inference operations.
Patent Information
- Application Number
- JP2024063157
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-04-11
- Filing Date
- 2024-04-10
- Publication Date
- 2024-11-13
AI Technical Summary
Existing methods for processing input quantities, particularly binary and ternary data, are limited in their ability to efficiently convert and process current signals through transistors, leading to inefficiencies in operations such as multiplication and accumulation.
A method and apparatus utilizing field effect transistors (FETs), including ferroelectric FETs (FeFETs), to selectively connect load sections of transistors with circuit nodes based on input quantities, enabling current-based analog-to-digital conversion and differential processing to encode and process ternary data, facilitating operations like multiply and accumulate (MAC) for neural network inference.
Enhances the processing of binary and ternary input quantities by improving current signal conversion and combination, allowing for efficient hardware-based operations like MAC, thereby accelerating neural network inference.
Smart Images

Figure 2024160206000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to a method for processing an input volume. The present disclosure further relates to an apparatus for processing an input quantity. Summary of the Invention [Means for solving the problem]
[0002] Exemplary embodiments relate to a method, such as a computer-implemented method, for processing an input quantity, comprising the steps of applying a first input quantity to a control electrode of a first transistor and applying a second input quantity to a current derivation mechanism, the current derivation mechanism being configured to selectively connect a load section of the first transistor with a first circuit node or a second circuit node based on the second input quantity, which in further exemplary embodiments allows for processing of the input quantity.
[0003] In a further exemplary embodiment, the method is contemplated to include a step of determining the output quantity based on at least one of the following factors: a) a quantity characteristic of a current flowing through the first circuit node; b) a quantity characteristic of a current flowing through the second circuit node.
[0004] In a further exemplary embodiment, it is contemplated that the method comprises determining a quantity characteristic of a current flowing through a first circuit node using a first, e.g., current-based, analog-to-digital conversion mechanism, and determining a quantity characteristic of a current flowing through a second circuit node using a second, e.g., current-based, analog-to-digital conversion mechanism.
[0005] In a further exemplary embodiment, it is contemplated that the method comprises determining a difference between a quantity characteristic of a current flowing through the first circuit node and a quantity characteristic of a current flowing through the second circuit node using a differential, e.g. current-based, analog-to-digital conversion scheme.
[0006] In a further exemplary embodiment, it is contemplated that the first input quantity is a binary quantity. In a further exemplary embodiment, it is contemplated that the second input quantity is a binary quantity, for example the second input quantity is characterized by a sign, for example "plus" or "minus". Thus, for example, ternary quantities can be processed in this way, which ternary quantities can take, for example, the three values "-1", "0", "1".
[0007] In a further exemplary embodiment, it is contemplated that the method comprises applying a first input quantity to a control electrode of a second transistor, in which a current directing mechanism is formed to connect the load section of the first transistor with a first circuit node and the load section of the second transistor with a second circuit node when the second input quantity has a first state, and the current directing mechanism is formed to connect the load section of the first transistor with the second circuit node and the load section of the second transistor with the first circuit node when the second input quantity has a second state.
[0008] In a further exemplary embodiment, it is contemplated that the method comprises using a field effect transistor FET, for example a MOSFET, as the first transistor and / or the one or second transistor.
[0009] In a further exemplary embodiment, it is contemplated that the method comprises using a ferroelectric field effect transistor FeFET as the first transistor and / or as the one or second transistor.
[0010] In a further exemplary embodiment, it is contemplated that the method comprises at least one of the following elements: a) adjusting, e.g., programming, a threshold voltage of at least a first transistor based on a first input quantity; b) adjusting the threshold voltage of the first transistor to a first, e.g., relatively low, value (e.g., for possible values of the first input quantity); and c) adjusting the threshold voltage of the second transistor to a second, e.g., relatively high, value (e.g., for possible values of the first input quantity).
[0011] In a further exemplary embodiment, it is contemplated that the method includes a step of characterizing, e.g., encoding, the numerical value of the ternary quantity using at least a first transistor and / or a first input quantity, and a step of characterizing, e.g., encoding, the sign of the ternary quantity (e.g., plus, "+" or minus, "-") using a second input quantity.
[0012] A further exemplary embodiment relates to an apparatus for carrying out the method according to the embodiment. In further exemplary embodiments, the apparatus has a first transistor, a current directing mechanism, and optionally at least one or a second transistor, e.g., it is contemplated that the apparatus has a control device for controlling at least one aspect of a method according to an embodiment.
[0013] In a further exemplary embodiment, it is contemplated that a first transistor and a second transistor are provided, and a first input quantity can be applied to control electrodes of the first transistor and the second transistor, for example, the control electrodes of the first transistor and the second transistor being interconnected.
[0014] In further exemplary embodiments, it is contemplated that a) a first terminal of the load section of the first transistor is connected to a first terminal of the current directing mechanism, and / or b) a second terminal of the load section of the first transistor is connected to a first reference potential, e.g., ground potential, and / or c) a first terminal of the load section of the second transistor is connected to the second terminal of the current directing mechanism, and / or d) a second terminal of the load section of the second transistor is connected to the first reference potential.
[0015] In a further exemplary embodiment, it is contemplated that the current derivation mechanism has a first transistor, a first terminal of which is selectively connectable to the second circuit node by the first transistor based on a second input quantity, and that the current derivation mechanism has a second transistor, a first terminal of which is selectively connectable to the first circuit node by the second transistor based on the second input quantity.
[0016] In a further exemplary embodiment, it is contemplated that the current derivation mechanism has a third transistor, a second terminal of which is selectively connectable to the second circuit node by the third transistor based on the second input quantity, and that the current derivation mechanism has a fourth transistor, a second terminal of which is selectively connectable to the first circuit node by the fourth transistor based on the second input quantity.
[0017] Further exemplary embodiments relate to a computing arrangement, such as a vector matrix multiplication arrangement, having at least one device according to the embodiment. A further exemplary embodiment relates to a computer-readable memory medium comprising instructions that, when executed by a computer, cause the computer to perform a method according to the embodiment.
[0018] A further exemplary embodiment relates to a computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method according to the embodiment. Further exemplary embodiments relate to a data carrier signal carrying and / or featuring a computer program according to the embodiment.
[0019] Further exemplary embodiments relate to the use of a method according to an embodiment and / or a device according to an embodiment and / or a computing mechanism according to an embodiment and / or a computer readable memory medium according to an embodiment and / or a computer program according to an embodiment and / or a data carrier signal according to an embodiment for at least one of the following aspects: a) coding, e.g. of weights, e.g. of ternary values, b) providing a circuit for performing, e.g. binomial or ternary multiplication and / or accumulation operations, e.g. multiply and accumulate (MAC), c) utilizing transistors, e.g. of FeFET type, for performing ternary multiplication and / or accumulation operations, e.g. multiply and accumulate (MAC), d) accelerating, e.g. hardware-based acceleration, the inference of a ternary artificial neural network.
[0020] Further features, applicability and advantages of the present invention will become apparent from the following description of exemplary embodiments of the invention as illustrated in the drawing figures, in which all described or illustrated features, either alone or in any combination, are the subject of the present invention, independently of their grouping in the claims or their relevant parts, and independently of their representation or display in the specification or drawings. [Brief description of the drawings]
[0021] [Figure 1] FIG. 1 illustrates a simplified flow diagram according to an exemplary embodiment. [Diagram 2] FIG. 1 illustrates a simplified block diagram according to an exemplary embodiment. [Diagram 3] FIG. 1 illustrates a simplified flow diagram according to an exemplary embodiment. [Figure 4]FIG. 1 illustrates a simplified flow diagram according to an exemplary embodiment. [Diagram 5] FIG. 1 illustrates a simplified circuit diagram according to an exemplary embodiment. [Figure 6] FIG. 1 illustrates a simplified flow diagram according to an exemplary embodiment. [Figure 7] FIG. 1 illustrates a simplified flow diagram according to an exemplary embodiment. [Figure 8] 8A and 8B are schematic diagrams of a simplified flow diagram according to an exemplary embodiment; [Figure 9] FIG. 1 illustrates a simplified circuit diagram according to an exemplary embodiment. [Figure 10] FIG. 1 illustrates a simplified block diagram according to an exemplary embodiment. [Figure 11] FIG. 13 shows a simplified block diagram according to a further exemplary embodiment. [Figure 12] 1A-1D are schematic diagrams illustrating modes of use according to exemplary embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0022] Exemplary embodiments (see Figs. 1, 2) relate to a method, for example a computer-implemented method, for processing input quantities, comprising a step 200 of applying a first input quantity E1 to a control electrode 110-1a (Fig. 2) of a first transistor 110-1 and a step 202 (Fig. 1) of applying a second input quantity E2 to a current derivation mechanism 120, which is formed for selectively connecting the load section 110-1-LS of the first transistor 110-1 with a first circuit node N1 or a second circuit node N2 based on the second input quantity E2, which in further exemplary embodiments allows processing of the input quantities E1, E2, for example for performing calculations which will be described in more detail exemplarily below.
[0023] For example, the current deriving mechanism 120 is connected to a first terminal (not shown in FIG. 2) of the load section 110-1-LS, and a second terminal of the load section 110-1-LS is connected to a reference potential BP-1, for example, ground potential, so that, for example, the current I-110-LS flowing through the load section 110-1-LS as required (for example, in the case of a corresponding control of the first transistor 110-1 by the first input quantity E1) can be selectively derived from the first circuit node N1 or the second circuit node N2 based on the second input signal E2 (see currents I-N1, I-N2 according to FIG. 2).
[0024] In a further exemplary embodiment, Fig. 1, it is contemplated that the method comprises a step 204 of determining an output quantity A1 on the basis of at least one of the following elements: a) a quantity GI-N1 characterizing the current I-N1 flowing through the first circuit node N1 (Fig. 2), b) a quantity GI-N2 characterizing the current I-N2 flowing through the second circuit node N2. In a further exemplary embodiment, the output quantity A1 or the value of the output quantity A1 depends on at least one of the following elements: a) a first input quantity E1, b) a second input quantity E2, c) a characteristic of the first transistor 110-1.
[0025] In further exemplary embodiments, the quantities GI-N1, GI-N2 may for example be directly the respective currents, for example the respective progressions of the currents over time. In further exemplary embodiments, the quantities GI-N1, GI-N2 may for example be voltages associated with the respective currents I-N1, I-N2.
[0026] In a further exemplary embodiment, Fig. 3, the method is contemplated to have a step 210 of determining a quantity GI-N1 characteristic of the current I-N1 flowing through the first circuit node N1, for example using a first, current-based (or current-driven) analog-to-digital converter ADC1 (Fig. 2), and a step 212 of determining a quantity GI-N2 characteristic of the current I-N2 flowing through the second circuit node N2, for example using a second, current-based analog-to-digital converter ADC2. Optional block 214 according to Fig. 3 symbolizes the determination of an output quantity A1, for example based on information obtained according to blocks 210 and 212.
[0027] In a further exemplary embodiment, Fig. 4, the method is provided for with a step 220 of determining a difference DIFF between a quantity GI-N1 characteristic of the current I-N1 flowing through the first circuit node N1 and a quantity GI-N2 characteristic of the current I-N2 flowing through the second circuit node N2, using a differential, for example current-based, analog-to-digital converter ADC'. Optional block 222 according to Fig. 4 symbolizes the determination of an output quantity A1, for example on the basis of the information DIFF obtained according to block 220.
[0028] In a further exemplary embodiment, it is contemplated that the first input quantity E1 is a binary quantity and may take, for example, a value 1 or 0 or a logical 1 or logical 0. Thereby, in a further exemplary embodiment, when the control electrode 110-1a is controlled based on the value "1" of the first input quantity, the load section 110-1-LS of the first transistor 110-1 becomes, for example, low resistance. Thereby, in a further exemplary embodiment, when the control electrode 110-1a is controlled based on the value "0" of the first input quantity, the load section 110-1-LS of the first transistor 110-1 becomes, for example, high resistance.
[0029] In a further exemplary embodiment, it is contemplated that the second input quantity E2 is a binary quantity, for example the second input quantity being characterized by a sign, for example "plus" or "minus". Thus, in this way, the device 100 according to Fig. 2 can process, for example, ternary quantities, which can take, for example, the three values "-1", "0", "1".
[0030] In further exemplary embodiments, the configuration 110-1, 120 according to Fig. 2 can be combined with, for example, further transistors and / or current derivation mechanisms (not shown), which also act on the first or second circuit node N1, N2 (e.g. based on a corresponding second input quantity to the further current derivation mechanism) to, for example, combine, for example, the currents of different transistors. Thus, in further exemplary embodiments, for example multiplication and / or accumulation operations, for example multiply and accumulate (MAC) operations, can be performed on a hardware basis.
[0031] In a further exemplary embodiment (see Figs. 5 and 6), two exemplary transistors 110-1, 110-2 are used in the device 100a, both of which are connected to a current directing mechanism 120a (Fig. 5). An exemplary method is symbolized in Fig. 6. For example, it is contemplated that the method includes a step 234 of applying a first input quantity E1 to the control electrode 110-2a (FIG. 5) of the second transistor 110-2, in which the current directing mechanism 120a is formed to connect the load section 110-1-LS of the first transistor 110-1 to the first circuit node N1 and the load section 110-2-LS of the second transistor 110-2 to the second circuit node N2 when the second input quantity E2 has a first state, and the current directing mechanism 120a is formed to connect the load section 110-1-LS of the first transistor 110-1 to the second circuit node N2 and the load section 110-2-LS of the second transistor 110-2 to the first circuit node N1 when the second input quantity E2 has a second state.
[0032] Optionally, the method according to FIG. 6 comprises further steps 230, 232, which correspond for example to steps 200, 202 according to FIG. Exemplarily, based on blocks 230, 232, 234, an output quantity A1' can be determined in block 236 based on FIG. 6, which output quantity A1' depends, for example, on input quantities E1, E2 and possibly on characteristics of transistors 110-1, 110-2 (e.g. threshold voltage, resistance in low resistance state, etc.), e.g. similar to the determination of output quantity A1 based on block 204 of FIG. 1.
[0033] In a further exemplary embodiment, FIG. 7, the method is contemplated to have a step 240 of using a field effect transistor FET, e.g. a MOSFET, as the first transistor 110-1 and / or the one or second transistor 110-2.
[0034] In a further exemplary embodiment, FIG. 7, the method is contemplated to have a step 242 of using a ferroelectric field effect transistor FeFET as the first transistor 110-1 and / or as one or the second transistor 110-2.
[0035] In a further exemplary embodiment, FIG. 8A, the method is contemplated to have at least one of the following elements: a) adjusting, e.g., programming 250, the threshold voltage UTH of at least the first transistor 110-1 (or, e.g., both transistors 110-1, 110-2) based on a first input quantity E1 (e.g., when the transistors 110-1, 110-2 are formed as FeFETs); b) adjusting 252 the threshold voltage of the first transistor 110-1 to a first, e.g., relatively low value W1 (e.g., for possible values of the first input quantity E1); c) adjusting 254 the threshold voltage of the second transistor 110-2 to a second, e.g., relatively high value W2 (e.g., for possible values of the first input quantity E1).
[0036] In a further exemplary embodiment, FIG. 8B, the method is contemplated to have a step 260 of characterizing, e.g. encoding, the numerical value BETR of the ternary quantity using at least the first transistor 110-1 and / or a first input quantity E1, and a step 262 of characterizing, e.g. encoding, the sign VZ of the ternary quantity (e.g. plus, “+” or minus, “-”) using a second input quantity E2.
[0037] Further exemplary embodiments, figures 2, 5, relate to an apparatus 100, 100a for carrying out the method according to the embodiment. In further exemplary embodiments, the device 100, 100a has a first transistor 110-1, a current directing mechanism 120, 120a, and optionally at least one or a second transistor 110-2, e.g., it is contemplated that the device 100, 100a has an optional control device 130 for controlling at least one aspect of a method according to the embodiment.
[0038] In a further exemplary embodiment, FIG. 5 , a first transistor 110-1 and a second transistor 110-2 are provided (e.g., each formed as an FeFET), and a first input quantity E1 can be applied to control electrodes 110-1a, 110-2a of the first transistor 110-1 and the second transistor 110-2, for example, the control electrodes 110-1a, 110-2a of the first transistor 110-1 and the second transistor 110-2 being contemplated to be interconnected.
[0039] In further exemplary embodiments, it is contemplated that a) the first terminal 110-1b of the load section 110-1-LS (e.g., drain-source section) of the first transistor 110-1 is connected to the first terminal 120a-1 of the current directing mechanism 120a, and / or b) the second terminal 110-1c of the load section 110-1-LS of the first transistor 110-1 is connected to one or a first reference potential, e.g., ground potential BP-1, and / or c) the first terminal 110-2b of the load section 110-2-LS of the second transistor 110-2 is connected to the second terminal 120a-2 of the current directing mechanism 120a, and / or d) the second terminal 110-2c of the load section 110-2-LS of the second transistor 110-2 is connected to the first reference potential BP-1. This allows the transistors 110-1, 110-2 to derive current from the circuit nodes N1, N2 to the ground potential BP-1 via the current deriving mechanism 120a as needed based on, among other things, the first input quantity E1 and the second input quantity E2.
[0040] In a further exemplary embodiment, FIG. 5, it is contemplated that the current derivation mechanism 120a has a first transistor 122a, a first terminal 120a-1 of the current derivation mechanism 120a is selectively connectable to a second circuit node N2 by the first transistor 122a based on a second input quantity E2, and that the current derivation mechanism 120a has a second transistor 122b, a first terminal 120a-1 of the current derivation mechanism 120a is selectively connectable to a first circuit node N1 by the second transistor 122b based on the second input quantity E2.
[0041] In a further exemplary embodiment, FIG. 5, it is contemplated that the current derivation mechanism 120a has a third transistor 122c, by which the second terminal 120a-2 of the current derivation mechanism 120a is selectively connectable to the second circuit node N2 based on the second input quantity E2, and that the current derivation mechanism 120a has a fourth transistor 122d, by which the second terminal 120a-2 of the current derivation mechanism 120a is selectively connectable to the first circuit node N1 based on the second input quantity E2.
[0042] In a further exemplary embodiment, the first transistor 122a of the current derivation mechanism 120a is, for example, a FET, for example of a PMOS type. In a further exemplary embodiment, the second transistor 122b of the current derivation mechanism 120a is, for example, a FET, for example of an NMOS type. In a further exemplary embodiment, the third transistor 122c of the current derivation mechanism 120a is, for example, a FET, for example of an NMOS type. In a further exemplary embodiment, the fourth transistor 122d of the current derivation mechanism 120a is, for example, a FET, for example of a PMOS type.
[0043] In a further exemplary embodiment, the first transistor 110-1 is, for example, an NMOS-based FeFET, as is the second transistor 110-2, for example. In a further exemplary embodiment, the device 100a according to Fig. 5 can be used for example to perform a MAC operation, which can be used for example to evaluate (inference) a neural network. For example, a ternary input feature can be coded by a first input quantity E1 (for example for a numerical value, e.g. 1 or 0) and a second input quantity E2 (for example for a sign "+" or "-"), in which case, for example, weights can be coded by the FeFETs 110-1, 110-2, for example by corresponding programming of the threshold voltages of the FeFETs.
[0044] 5, both FeFETs are programmed to a relatively low threshold voltage (LVT) when they are to be characterized, e.g., store, a logical "1" value. In this situation, FeFETs 110-1, 110-2 may be activated (transition to a low resistance state) when, e.g., a non-zero gate voltage is applied to their gate electrodes 110-1a, 110-2a.
[0045] 5, both FeFETs are programmed, for example, to a relatively high threshold voltage (HVT) when they are to be characterized, for example, as a logical "0" value, e.g., when they are to store a logical "0" value. In this situation, FeFETs 110-1, 110-2 cannot be activated, for example, when a non-zero gate voltage is applied.
[0046] In a further exemplary embodiment, a "1" is encoded, for example, by programming a first FeFET of both FeFETs, here for example the first FeFET 110-1, in LVT, i.e., to a relatively low threshold voltage, and a second FeFET of both FeFETs, here for example the second FeFET 110-2, in HVT, i.e., to a relatively high threshold voltage.
[0047] In a further exemplary embodiment, assuming, for example, that a value "logical 1" is encoded by programming FeFET 110-1 to the LVT and programming FeFET 110-2 to the HVT, a value of, for example, "-1" may be encoded by inverting the state of both FeFETs 110-1, 110-2, i.e., by programming FeFET 110-1 to the HVT and FeFET 110-2 to the LVT.
[0048] As already explained above, the current derivation mechanism 120a in an exemplary embodiment allows selective connection of the load sections 110-1-LS, 110-2-LS of the FeFETs 110-1, 110-2 with the circuit nodes N1, N2 based on the second input quantity E2, so that in a further exemplary embodiment, the first circuit nodes of both circuit nodes N1, N2 can "collect" or combine (by adding up the respective currents) the processing or calculation results associated with, for example, positive values, e.g., "+1", and the second circuit nodes of both circuit nodes N1, N2 can "collect" or combine (by adding up the respective currents) the processing or calculation results associated with, for example, negative values, e.g., "-1".
[0049] In further embodiments, for example, more than the two FeFETs 110-1, 110-2 illustratively described above can be provided, with the further FeFETs being able to connect their respective load sections to circuit nodes N1, N2 in a manner equivalent to FIG. 5, for example via corresponding further current directing mechanisms (not shown, see the vertical dots on the right side of FIG. 5), thereby summing the respective currents at circuit nodes N1, N2.
[0050] In a further exemplary embodiment, a differential analog-to-digital converter ADC' is preferably used to capture the currents I-N1, I-N2 through the circuit nodes N1, N2, since this eliminates leakage currents, e.g., due to temporarily deactivated FeFETs, and therefore the output quantities A1, A1' are not harmed by this.
[0051] In further exemplary embodiments, the FeFETs 110-1, 110-2 may be programmed with additional (e.g., different) values for their respective threshold voltages, providing additional flexibility in processing input quantities. Illustratively, such configurations may code values such as "-2", "-1", "0", "1", and "2".
[0052] In a further exemplary embodiment, FIG. 9, at least one transistor 110-1 (see also FIG. 2, FIG. 5) can be assigned a memristive element 12b, which can affect a charging current IL for the gate electrode 110-1a of the transistor 110-1, for example, based on its memory value or resistance value. FIG. 9 additionally shows, by way of example, a voltage source 12a, which provides a charging voltage V-1 for the gate electrode 110-1a of the transistor 110-1, which can be provided via the memristive element 12b for the gate electrode 110-1a of the transistor 110-1, as shown. The configuration 12 exemplarily shown in FIG. 9 can be used in further exemplary embodiments for one or more transistors 110-1, 110-2, ... of the device 100, 100a, for example, for providing or affecting a first input quantity E1 (FIG. 2).
[0053] FIG. 10 illustrates generally a simplified block diagram of an arrangement 300 according to an exemplary embodiment, using which, for example, optional control device 130 (FIGS. 2, 5) can be implemented.
[0054] The arrangement or apparatus 300 comprises, for example, a computing mechanism ("computer") 302 with at least one computing core and a memory mechanism 304 allocated to the computing mechanism 302 for at least temporarily storing at least one of the following elements: a) data DAT (e.g. data associated with a first input quantity E1 and / or a second input quantity E2 and / or an output quantity A1, A1'); b) a computer program PRG, for example for executing a method according to an embodiment.
[0055] In a further exemplary embodiment, the memory mechanism 304 comprises volatile memory (e.g., main memory (RAM)) 304a, and / or non-volatile (NVM) memory (e.g., flash EEPROM) 304b, or a combination thereof or with other memory types not expressly listed.
[0056] A further exemplary embodiment, FIG. 10, relates to a computer-readable memory medium SM comprising instructions PRG′ which, when executed by a computer 302, cause the computer 302 to carry out a method according to the embodiment.
[0057] A further exemplary embodiment relates to a computer program PRG, PRG' comprising instructions which, when executed by a computer 302, cause the computer 302 to carry out a method according to an embodiment.
[0058] Further exemplary embodiments relate to a data carrier signal DCS which is characteristic of and / or carries a computer program PRG, PRG' according to an embodiment. The data carrier signal DCS can for example be received via the optional data interface 306 of the device 300.
[0059] In further exemplary embodiments, the apparatus 300 may control, e.g., at least temporarily, at least one of the following aspects: a) providing a first input quantity E1; b) providing a second input quantity E2; c) determining the output quantities A1, A1'; d) programming of an optional FeFET (e.g., the threshold voltage of the FeFET); and e) programming of an optional memristive element 12b.
[0060] A further exemplary embodiment, Fig. 11, relates to a computational arrangement, for example a vector matrix multiplication arrangement VMM, having at least one device 100, 100a according to an embodiment. In a further exemplary embodiment, the first and / or second input quantities E1, E2 can for example be assigned to an aspect or sub-aspect of a vector matrix multiplication, which aspect or sub-aspect of a vector matrix multiplication can be performed using the device 100, 100a (possibly supplemented with a corresponding number of further transistors and / or current derivation mechanisms), for example on a hardware basis.
[0061] A further exemplary embodiment, FIG. 12, relates to the use 400 of a method according to an embodiment and / or a device according to an embodiment 100, 100a and / or a computing mechanism VMM according to an embodiment and / or a computer readable memory medium SM according to an embodiment and / or a computer program PRG, PRG' according to an embodiment and / or a data carrier signal DCS according to an embodiment for at least one of the following aspects: a) coding 401, e.g. of ternary values, e.g. of weights, e.g. for an artificial neural network; b) providing a circuit for performing 402, e.g. binomial or ternary multiplication and / or accumulation operations, e.g. multiply and accumulate (MAC); c) using transistors, e.g. of FeFET type, for performing ternary multiplication and / or accumulation operations, e.g. multiply and accumulate (MAC) 403; d) accelerating 404, e.g. hardware-based acceleration, the inference of, e.g. a ternary artificial neural network.
[0062] Illustratively, the apparatus 100a based on FIG. 5 can be used to perform, for example, the following calculation: 1*1=1, where "*" is the scalar multiplication operator, 0*1=0, 1*0=0, 0*-1=0, -1*0=0, 1*-1=-1, -1*1=-1, -1*-1=1.
[0063] The project leading this application has been funded by the ECSEL Collaboration (JU) under grant agreement no. 826655. The Collaboration receives support from the European Union's Horizon 2020 research and innovation programme and from Belgium, France, Germany, the Netherlands and Switzerland. [Explanation of symbols]
[0064] 100;100a Equipment 110-1 First transistor 110-1a: control electrode of first transistor 110-1b First terminal of the load section of the first transistor 110-1c second terminal of the load section of the first transistor 110-1-LS Load section of the first transistor 110-2 Second transistor 110-2a: control electrode of second transistor 110-2b: first terminal of the load section of the second transistor 110-2c Second terminal of the load section of the second transistor 110-2-LS Second transistor load section 120;120a Current derivation mechanism 120a-1 First terminal of current lead-out mechanism 120a-2 Second terminal of current lead-out mechanism 122a First transistor of current directing mechanism 122b second transistor of current directing mechanism 122c Third transistor of current directing mechanism 122d Fourth transistor of current directing mechanism 130;300 Control device 300 Configuration or Equipment 302 Computer 304 Memory Mechanism 306 Data Interface A1;A1' Output amount ADC1 First analog / digital converter ADC2 Second analog / digital converter ADC' differential analog / digital conversion mechanism BETR ternary amount value BP-1 First reference potential DCS Data Carrier Signal DIFF The difference between a quantity characteristic of the current flowing through a first circuit node and a quantity characteristic of the current flowing through a second circuit node. E1 First input quantity E2 Second input quantity GI-N1 A quantity characterizing the current flowing through the first circuit node GI-N2 A quantity characterizing the current flowing through the second circuit node I-N1 Current flowing through the first circuit node I-N2 Current flowing through the second circuit node N1 First circuit node N2 Second circuit node PRG Computer Program PRG' instruction SM Memory Media UTH Threshold Voltage VMM Vector Matrix Multiplication Mechanism VZ Sign of ternary quantity W1 First value W2 Second value
Claims
1. 1. A method, for example a computer-implemented method, for processing an input quantity, comprising the steps (200; 230) of applying a first input quantity (E1) to a control electrode (110-1a) of a first transistor (110-1) and the steps (202; 232) of applying a second input quantity (E2) to a current directing mechanism (120), the current directing mechanism (120) being configured to selectively connect a load section (110-1-LS) of the first transistor (110-1) to a first circuit node (N1) or a second circuit node (N2) based on the second input quantity (E2).
2. 2. The method of claim 1, comprising determining (204) an output quantity (A1) based on at least one of the following factors: a) a quantity (G-I-N1) characteristic of a current (I-N1) flowing through said first circuit node (N1); b) a quantity (G-I-N2) characteristic of a current (I-N2) flowing through said second circuit node (N2).
3. 3. The method of claim 2, further comprising determining (210) the quantity (G-I-N1) characteristic of the current (I-N1) flowing through the first circuit node (N1) using a first, e.g., current-based, analog-to-digital converter (ADC1), and determining (212) the quantity (G-I-N2) characteristic of the current (I-N2) flowing through the second circuit node (N2) using a second, e.g., current-based, analog-to-digital converter (ADC2).
4. 4. The method of claim 2 or 3, comprising determining (220) a difference (DIFF) between the quantity (G-I-N1) characterizing the current (I-N1) flowing through the first circuit node (N1) and the quantity (G-I-N2) characterizing the current (I-N2) flowing through the second circuit node (N2) using a differential, e.g. current-based, analog-to-digital converter (ADC').
5. The method according to any one of claims 1 to 4, wherein the first input quantity (E1) is a binary quantity.
6. 6. The method according to claim 1, wherein the second input quantity (E2) is a binary quantity, for example the second input quantity (E2) is characterized by a sign, for example "plus" or "minus".
7. and applying (234) the first input quantity (E1) to a control electrode (110-2a) of a second transistor (110-2), in which the current directing mechanism (120a) couples the load section (110-1-LS) of the first transistor (110-1) to the first circuit node (N1) and the load section (110-2-LS) of the second transistor (110-2) to the first circuit node (N2) when the second input quantity (E2) has a first state.
7. The method of claim 6, wherein the current directing mechanism (120a) is formed to connect the load section (110-1-LS) of the first transistor (110-1) to the second circuit node (N2) and the load section (110-2-LS) of the second transistor (110-2) to the first circuit node (N1) when the second input quantity (E2) has a second state.
8. The method according to any one of the preceding claims, comprising the step (240) of using a field effect transistor FET as the first transistor (110-1) and / or as one or the second transistor (110-2).
9. The method according to any one of the preceding claims, comprising the step (242) of using a ferroelectric field effect transistor (FeFET) as the first transistor (110-1) and / or as one or the second transistor (110-2).
10. 10. The method of claim 9, comprising at least one of the following elements: a) adjusting (250), e.g. programming, a threshold voltage (UTH) of at least the first transistor (110-1) based on the first input quantity (E1); b) adjusting (252) the threshold voltage of the first transistor (110-1) to a first, e.g. relatively low, value (W1); c) adjusting (254) the threshold voltage of the second transistor (110-2) to a second, e.g. relatively high, value (W2).
11. The method according to any one of claims 1 to 10, comprising a step (260) of characterising, e.g. encoding, a numerical value (BETR) of a ternary quantity using at least the first transistor (110-1) and / or the first input quantity (E1), and a step (262) of characterising, e.g. encoding, a sign (VZ) of the ternary quantity using the second input quantity (E2).
12. Apparatus (100; 100a) for carrying out the method according to any one of the claims 1 to 11.
13. An apparatus (100; 100a) according to claim 12, comprising the first transistor (110-1), the current directing mechanism (120; 120a), and optionally at least one or the second transistor (110-2), e.g. the apparatus (100; 100a) comprising a control device (130; 300) for controlling at least one aspect of the method according to any one of claims 1 to 11.
14. The device (100a) according to claim 12 or 13, wherein the first transistor (110-1) and the second transistor (110-2) are provided, and the first input quantity (E1) can be applied to control electrodes (110-1a, 110-2a) of the first transistor (110-1) and the second transistor (110-2), for example the control electrodes (110-1a, 110-2a) of the first transistor (110-1) and the second transistor (110-2) are interconnected.
15. a) a first terminal (110-1b) of the load section (110-1-LS) of the first transistor (110-1) is connected to a first terminal (120a-1) of the current directing mechanism (120a); and / or b) a second terminal (110-1c) of the load section (110-1-LS) of the first transistor (110-1) is connected to a first reference potential (BP-1), e.g., ground potential; and / or c) 15. The apparatus (100a) of claim 14, wherein a first terminal (110-2b) of the load section (110-2-LS) of the second transistor (110-2) is connected to a second terminal (120a-2) of the current directing mechanism (120a), and / or d) a second terminal (110-2c) of the load section (110-2-LS) of the second transistor (110-2) is connected to the first reference potential (BP-1).
16. 16. The apparatus (100a) of claim 15, wherein the current directing mechanism (120a) comprises a first transistor (122a), the first terminal (120a-1) of the current directing mechanism (120a) is selectively connectable to the second circuit node (N2) by the first transistor (122a) based on the second input quantity (E2), and the current directing mechanism (120a) comprises a second transistor (122b), the first terminal (120a-1) of the current directing mechanism (120a) is selectively connectable to the first circuit node (N1) by the second transistor (122b) based on the second input quantity (E2).
17. 17. The apparatus (100a) of claim 16, wherein the current directing mechanism (120a) comprises a third transistor (122c), the second terminal (120a-2) of the current directing mechanism (120a) being selectively connectable to the second circuit node (N2) by the third transistor (122c) based on the second input quantity (E2), and wherein the current directing mechanism (120a) comprises a fourth transistor (122d), the second terminal (120a-2) of the current directing mechanism (120a) being selectively connectable to the first circuit node (N1) by the fourth transistor (122d) based on the second input quantity (E2).
18. A computing arrangement, for example a Vector Matrix Multiplication Machine (VMM), comprising at least one device (100; 100a) according to any one of claims 12 to 17.
19. A computer readable memory medium (SM) comprising instructions (PRG') which, when executed by a computer (302), cause said computer (302) to perform the method according to any one of claims 1 to 11.
20. A computer program (PRG) comprising instructions which, when executed by a computer (302), cause said computer (302) to carry out the method according to any one of claims 1 to 11.
21. A data carrier signal (DCS) carrying and / or being characterized by a computer program (PRG) according to claim 20.
22. The following aspects: a) Coding (401), e.g., weights, e.g., of ternary values; b) providing circuitry (402) for performing, e.g., binomial or ternary multiplication and / or accumulation operations, e.g., multiply and accumulate (MAC); c) the use of transistors, e.g. of FeFET type, to perform ternary multiplication and / or accumulation operations, e.g. multiply and accumulate (MAC) (403); d) Acceleration of the inference of the ternary artificial neural network (404), e.g., hardware-based acceleration; Use (400) of a method according to any one of claims 1 to 11 and / or an apparatus (100; 100a) according to any one of claims 12 to 17 and / or a computing mechanism (VMM) according to claim 18 and / or a computer readable memory medium (SM) according to claim 19 and / or a computer program (PRG) according to claim 20 and / or a data carrier signal (DCS) according to claim 21 for at least one of
