Apparatus and method for in-memory binary convolution for accelerating deep binary neural networks based on a non-volatile memory structure
The present disclosure relates to a structure including a differential memory array circuit which is configured to perform a binary convolution of two input word operands by accumulating a summation of currents through a plurality of bits which are each arranged between a wordline and a sourceline in a horizontal direction and bitlines in a vertical direction.
1. A structure comprising a differential memory array circuit which is configured to perform a binary convolution of two input word operands by accumulating a summation of currents through a plurality of bits which are each arranged between a wordline and a sourceline in a horizontal direction and bitlines in a vertical direction,
wherein each of the plurality of bits comprise two devices,
a first device of the two devices comprises a source connected to a bitline and a drain connected to a first resistor of two resistors, and
a second device of the two devices comprises the source connected to a second resistor of the two resistors and a drain connected to a complementary bitline.
2. The structure of claim 1 , wherein the two devices are between the wordline and the sourceline in the horizontal direction and bitlines in the vertical direction.
3. The structure of claim 2 , wherein each of the two devices are a NMOS transistor.
4. The structure of claim 3 , wherein each of the plurality of bits comprise the two resistors.
5. The structure of claim 4 , wherein the first device of the two devices comprises a gate connected to the wordline and the second device of the two devices comprises the gate connected to the wordline.
6. The structure of claim 1 , wherein the sourceline is an input to an operational amplifier (OpAmp) and analog to digital converter (ADC).
7. The structure of claim 1 , wherein the differential memory array circuit further comprises a sense amplifier between the bitlines in the vertical direction.
8. The structure of claim 1 , wherein the differential memory array structure comprises a spin-transfer torque magnetic random-access memory (STT MRAM).
9. The structure of claim 1 , wherein the differential memory array structure comprises one of memristor devices and phase-change material devices.
10. The structure of claim 1 , further comprising an enhanced decoder configured to receive a first address to a first input word operand of the two input word operands and a second address to a second input word operand of the two input word operands, and the enhanced decoder further comprises a first address decoder configured to receive the first address to the first input word operand and a second address decoder configured to receive the second address to the second input word operand.
11. The structure of claim 10 , wherein the first address decoder is configured to output a first decoder output corresponding to the first address of the first input word operand to a writeline/sourceline driver and the second address decoder is configured to output a second decoder output corresponding to the second address of the second input word operand to the writeline/sourceline driver.
12. A circuit, comprising:
an enhanced decoder which is configured to receive a first address to the first input word operand and a second address to the second input word operand and output a binary convolution output of the first input word operand and the second input word operand stored in a memory; and
an operational amplifier (OpAmp) and analog to digital converter (ADC) configured to receive the binary convolution output and convert the binary convolution output to a digital output for an in-memory binary neural network (BNN) computation,
wherein the enhanced decoder further comprises a first address decoder configured to receive the first address to the first input word operand and a second address decoder configured to receive the second address to the second input word operand, and
the first address decoder is configured to output a first decoder output corresponding to the first address of the first input word operand to a writeline/sourceline driver and the second address decoder configured to output a second decoder output corresponding to the second address of the second input word operand to the writeline/sourceline driver.
13. The circuit of claim 12 , wherein the writeline/sourceline driver comprises an OR gate which is configured to receive the first decoder output and the second decoder output to drive a wordline signal.
14. The circuit of claim 13 , wherein the writeline/sourceline driver further comprises an inverter and four transistors to drive a sourceline signal.
15. The circuit of claim 14 , wherein the four transistors comprise two PMOS transistors and two NMOS transistors.
16. The circuit of claim 14 , wherein the writeline/sourceline driver further comprises another transistor which is configured to connect the sourceline signal to a compute-line signal in response to a binary neural network (BNN) enable signal being enabled.
17. The circuit of claim 14 , wherein the writeline/sourceline driver further comprises another transistor which is configured to connect the sourceline signal to ground in response to a binary neural network (BNN) enable signal not being enabled.
18. A method, comprising:
enabling a binary neural network (BNN) enable signal to perform an in-memory binary neural network (BNN) computation in a differential memory array circuit;
inputting a first address to the first input word operand and a second address to the second input word operand to the differential memory array circuit;
performing a binary convolution of the first input word operand and the second input word operand in the differential memory array circuit by accumulating a summation of currents through a plurality of bits in the differential memory array circuit; and
outputting a binary convolution output as a result of the summation of currents,
wherein, in response to the BNN enable signal being enabled, a first wordline is enabled, a NMOS transistor is turned off, and a first sourceline is connected to an operational amplifier (OpAmp) and analog to digital converter (ADC) through a compute-line CL.
19. The method of claim 18 , further comprising converting the binary convolution output to a digital output for an in-memory binary neural network (BNN) computation.
20. The method of claim 18 , further comprising:
enabling a first wordline signal and connecting a first sourceline signal to a voltage power supply signal in the differential memory array circuit; and
enabling a second wordline signal and connecting a second sourceline signal to a compute-line signal in the differential memory array circuit.