Switched capacitor vector-matrix multiplier
Methods and apparatuses enable a general-purpose low power analog vector-matrix multiplier. A switched capacitor matrix multiplier may comprise a plurality of successive approximate registers (SAR) operating in parallel, each SAR having a SAR digital output; and a plurality of Analog Multiply-and-Accumulate (MAC) units for multiplying and accumulating and scaling bit-wise products of a digital weight matrix with a digital input vector, wherein each MAC unit is connected in series to a SAR of the plurality of SARs.
1. A system configured to perform in-memory multiplication of a first multi-bit binary number with a second multi-bit binary number to generate a product of the first and second binary numbers, the system comprising:
a plurality of multiplier circuits connected in parallel, each multiplier circuit associated with different bits of the first and second binary numbers, wherein each multiplier circuit comprises one or more addressable memory elements storing one or more bits of an associated second binary number, and a unit capacitor, wherein each multiplier circuit is configured to multiply associated bits of the first and second binary numbers; and
a shared output to which the unit capacitor of each multiplier circuit is connected, wherein respective charges of the unit capacitor of each multiplier circuit are accumulated at the shared output, wherein the accumulated charge is representative of the product of the first and second binary numbers.
2. The system of claim 1 , wherein the one or more memory elements comprise one or more of an SRAM bit cell or a cross-coupled inverter.
3. The system of claim 1 , wherein each multiplier circuit comprises one or more of an AND gate, an inverter, or a multiplexer.
4. The system of claim 1 , wherein each multiplier circuit comprises a multiplexer having an output that drives the unit capacitor.
5. The system of claim 1 , wherein the plurality of multiplier circuits and the shared output are configured to perform one or more of matrix multiplication, matrix multiplier quantization, or scaling.
6. The system of claim 1 , wherein the system is configured to simultaneously and locally perform memory access and computation operations.
7. The system of claim 1 , wherein the plurality of multiplier circuits and the shared output are configured to perform quantization only once per multiplication operation independent of a resolution of the multiplication operation.
8. The system of claim 7 , wherein the multiplication operation comprises a multiply-and-accumulate (MAC) operation.
9. A method for performing in-memory multiplication of a first multi-bit binary number with a second multi-bit binary number to generate a product of the first and second binary numbers, the method comprising:
multiplying each bit of the first binary number with an associated bit of the second binary numbers via a different one of a plurality of multiplier circuits to generate a plurality of multiplication values, wherein the plurality of multiplier circuits are connected in parallel and wherein each multiplier circuit comprises one or more addressable memory elements storing one or more bits of an associated second binary number;
converting each multiplication value to a charge value by a unit capacitor associated with the multiplier circuit; and
accumulating charge values of the plurality of multiplier circuits at a shared output, wherein the accumulated charge is representative of the product of the first and second binary numbers.
10. The method of claim 9 , wherein each multiplier circuit comprises one or more memory elements.
11. The method of claim 10 , wherein the one or more memory elements comprise one or more of an SRAM bit cell or a cross-coupled inverter.
12. The method of claim 10 , wherein each multiplier circuit comprises one or more of an AND gate, an inverter, or a multiplexer.
13. The method of claim 10 , wherein each multiplier circuit comprises a multiplexer having an output that drives the unit capacitor.
14. The method of claim 9 , further comprising:
performing matrix multiplication using the plurality of multiplier circuits and the shared output.
15. The method of claim 9 , further comprising:
performing matrix multiplier quantization using the plurality of multiplier circuits and the shared output.
16. The method of claim 9 , further comprising:
performing scaling using the plurality of multiplier circuits and the shared output.
17. The method of claim 9 , further comprising:
simultaneously and locally performing memory access and computation operations using the plurality of multiplier circuits and the shared output.
18. The method of claim 9 , further comprising:
performing quantization, using the plurality of multiplier circuits and the shared output, only once per multiplication operation independent of a resolution of the multiplication operation.