Power efficient near memory analog multiply-and-accumulate (MAC)
A near memory system is provided for the calculation of a layer in a machine learning application. The near memory system includes an array of memory cells for storing an array of filter weights. A multiply-and-accumulate circuit couples to columns of the array to form the calculation of the layer.
1. A near memory system, comprising:
an array of memory cells arranged into a plurality of rows and into a plurality of columns, wherein each row is configured to store a plurality of filter weights for a layer in a machine-learning application;
a multiply-and-accumulate (MAC) circuit coupled to each column, the MAC circuit being configured to multiply the plurality of filter weights from an accessed-one of the rows with a plurality of inputs for the layer to provide a plurality of products, the MAC circuit being further configured to sum the plurality of products to provide an output for the layer;
a plurality of word lines corresponding to the plurality of rows, each word line extending across its corresponding row; and
a controller configured to sequentially assert the word lines to form the accessed-one of the rows.
2. The near memory system of claim 1 , wherein each memory cell is selected from the group consisting of a static random access memory (SRAM) bitcell, a magnetic RAM (MRAM) bitcell, a resistive RAM (RRAM) bitcell, a ferroelectric RAM (FRAM) bitcell, and a phase-change memory (PCM) bitcell.
3. The near memory system of claim 2 , wherein each SRAM bitcell is a six-transistor SRAM bitcell.
4. The near memory system of claim 1 , wherein the MAC circuit includes a plurality of logic gates corresponding to the plurality of columns, and wherein each logic gate is configured to multiply a filter weight with an input to form a binary product.
5. The near memory system of claim 4 , wherein each logic gate comprises an exclusive-not-or (XNOR) gate.
6. The near memory system of claim 5 , wherein the MAC circuit further includes a digital adder to add an output from each XNOR gate.
7. The near memory system of claim 5 , wherein the MAC circuit further includes:
a plurality of capacitors corresponding to the plurality of XNOR gates, and wherein each XNOR gate is configured to drive a first plate for the corresponding capacitor.
8. The near memory system of claim 7 , wherein the MAC circuit further includes an analog-to-digital converter coupled to a second plate for each capacitor.
9. The near memory system of claim 7 , wherein each input is a multi-bit input, the MAC circuit further comprising a sequential accumulator coupled to a second plate of each capacitor, wherein the sequential accumulator is configured to sequentially accumulate a charge from each second plate from a serial processing of the multi-bit inputs to provide an accumulation result.
10. The near memory system of claim 9 , wherein the layer is a fully-connected layer, and wherein the MAC circuit further includes an analog-to-digital converter configured to convert the accumulation result into an output for fully-connected layer.
11. A near memory system:
a plurality of arrays of memory cells, wherein each array is arranged into a plurality of rows and columns, and wherein each row is configured to store a plurality of filter weights for a layer in a machine-learning application, and wherein each array of memory cells is subdivided into a plurality of sub-arrays; and
a plurality of multiply-and-accumulate (MAC) circuits corresponding to the plurality of sub-arrays, each MAC circuit being coupled to each column in the corresponding sub-array, each MAC circuit being configured to multiply the plurality of filter weights from an accessed-one of the rows in the corresponding sub-array with a corresponding plurality of inputs to provide a plurality of products, the MAC circuit being further configured to sum the plurality of products to provide a sum, and wherein each array further includes an adder to add the sums from the array's MAC circuits to form an output for the array,
wherein each MAC circuit includes a plurality of logic gates corresponding to the plurality of columns in the corresponding sub-array, and wherein each logic gate is configured to multiply a filter weight with the corresponding one of the inputs to form a binary product.
12. The near memory system of claim 11 , wherein each memory cell is a static random access memory (SRAM) bitcell.
13. The near memory system of claim 12 , wherein each SRAM bitcell is a six-transistor SRAM bitcell.
14. The near memory system of claim 11 , wherein each logic gate comprises an exclusive-not-or (XNOR) gate.
15. The near memory system of claim 14 , wherein each MAC circuit further includes a digital adder to add an output from each of its XNOR gates.
16. The near memory system of claim 14 , wherein each MAC circuit further includes:
a plurality of capacitors corresponding to the plurality of XNOR gates in the MAC circuit, and wherein each XNOR gate is configured to drive a first plate for the corresponding capacitor.
17. Near memory system of claim 16 , wherein each MAC circuit further includes an analog-to-digital converter coupled to a second plate for each capacitor in the MAC circuit.
18. The near memory system of claim 16 , wherein each input is a multi-bit input, each MAC circuit further comprising a sequential accumulator coupled to a second plate of each of the MAC circuit's capacitors, wherein the MAC circuit's sequential accumulator is configured to sequentially accumulate a charge from each second plate for each capacitor in the MAC circuit for a serial processing of the multi-bit inputs to provide an accumulation result.
19. The near memory system of claim 18 , wherein the layer is a fully-connected layer, and wherein each MAC circuit further includes an analog-to-digital converter configured to convert the accumulation result into a digital value, and wherein each adder is further configured to add the digital value from the array's MAC circuits to provide an output for the fully-connected layer.
20. The near memory system of claim 11 , wherein the near memory system is integrated into a mobile device.
21. The near memory system of claim 20 , wherein the mobile device is cellular telephone.
22. A method for a near memory system, comprising:
asserting a word line for a row of memory cells to retrieve a plurality of filter weights for an output neuron in a layer for a machine learning application;
in a multiply-and-accumulate circuit coupled to the row of memory cells, multiplying the plurality of filter weights with a plurality of input neurons to form a plurality of products;
summing the plurality of products to calculate a value for the output neuron; and
sequentially asserting additional word lines to sequentially calculate values for additional output neurons in the layer.
23. The method of claim 22 , wherein summing the plurality of products comprises digitizing a voltage from a plurality of capacitors.
24. The method of claim 22 , wherein summing the plurality of products comprises a digital summing.