Binary weighted voltage encoding scheme for supporting multi-bit input precision
An illustrative embodiment disclosed herein is an apparatus including a non-volatile memory cell and multi-bit input circuitry that simultaneously receives a plurality of bits, receives a supply voltage, converts the plurality of bits and the supply voltage into a multiply voltage, and applies the multiply voltage to the non-volatile memory cell. The non-volatile memory cell may pass a memory cell current in response to the multiply voltage. A magnitude of the multiply voltage may represent a multiplier. The memory cell current may represent a product of the multiplier and a multiplicand stored in the non-volatile memory cell.
1. A non-volatile memory device, comprising:
a plurality of N input lines;
one or more conductive lines, including a first conductive line;
a plurality of memory cells, including a first plurality of N memory cells each connected between the first conductive line and a corresponding one of the input lines and each configured to store a first weight value of a neural network as a binary value; and
one or more control circuits connected to the input lines and the conductive lines and configured to:
receive an N-bit input value for the neural network;
convert the input value into a plurality of N voltage values;
concurrently apply each of the N voltage values to a corresponding one of the input lines; and
determine a resultant current in the first conductive line in response to concurrently applying each of the N voltages to the corresponding one of the input lines.
2. The non-volatile memory device of claim 1 , wherein:
the first weight value is a first bit of a multi-bit weight value,
the one or more conductive lines further includes a second conductive line,
the plurality of memory cells further includes a second plurality of N memory cells each connected between the second conductive line and a corresponding one of the input lines and each configured to store a second weight value, corresponding to a second bit of the multi-bit weight value as a binary value, and
wherein the one or more control circuits are further configured to:
determine a resultant current in the second conductive line in response to concurrently apply each of the N voltages to the corresponding one of the input lines concurrently with determining the resultant current in the first conductive line.
3. The non-volatile memory device of claim 2 , wherein the one or more control circuits are further configured to:
form a weighted sum of the resultant current in the first conductive line and the resultant current in the second conductive line; and
determine, from the weighted sum, a product of the input value for the neural network and the first weight value of the neural network.
4. The non-volatile memory device of claim 3 , wherein, in forming the weighted sum, the one or more control circuits are further configured to:
weight the resultant current in the first conductive line and the resultant current in the second conductive line according to a relative significance of the first bit of the multi-bit weight value and the second bit of the multi-bit weight value.
5. The non-volatile memory device of claim 1 , wherein, in converting the input value into the plurality of N voltage values, the one or more control circuits are further configured to:
receive a supply voltage; and
multiply a value of the supply voltage by each of N multiplier values to thereby generate the N voltage values each corresponding to a bit of the input value, the values of the N multiplier values corresponding to a relative significance of the corresponding bit.
6. The non-volatile memory device of claim 5 , wherein, in converting the input value into the plurality of N voltage values, the one or more control circuits are further configured to:
receive an additional voltage level; and
add the additional voltage level to each of the N voltages values prior to applying the N voltages to the corresponding input lines.
7. The non-volatile memory device of claim 1 , wherein the memory cells are magnetoresistive memory (MRAM) based memory cells.
8. The non-volatile memory device of claim 1 , wherein the memory cells are phase change memory (PCM) based memory cells.
9. A method, comprising:
receiving an N-bit input value for a neural network;
converting the input value into a plurality of N voltage values, each of the voltage values corresponding to a bit of the input value;
concurrently applying each of the N voltages to a corresponding one of a plurality of N input lines of an array of memory cells, the array including a first set of N memory cells each connected between a first conductive line and a corresponding one of the input lines, each memory cell of the first set storing a first weight value of the neural network as a binary value; and
determining a resultant current in the first conductive line in response to concurrently applying the N voltage values to the corresponding input lines.
10. The method of claim 9 , wherein converting the input value into the plurality of N voltage values comprises:
receiving a supply voltage; and
multiplying a value of the supply voltage by each of N multiplier values to thereby generate the N voltage values each corresponding to a bit of the input value, the values of the N multiplier values corresponding to a relative significance of the corresponding bit.
11. The method of claim 10 , wherein converting the input value into the plurality of N voltage values further comprises:
receiving an additional voltage level; and
adding the additional voltage level to each of the N voltages values prior to applying the N voltages to the corresponding input lines.
12. The method of claim 9 , wherein the first weight value is a first bit of a multi-bit weight value, the memory further including a second set of N memory cells each connected between a second conductive line and a corresponding one of the input lines, each memory cell of the second set storing a second bit of the multi-bit weight value as a binary value, the method further comprising:
determining a resultant current in the second conductive line in response to concurrently apply each of the N voltages to the corresponding one of the input lines concurrently with determining the resultant current in the first conductive line.
13. The method of claim 12 , further comprising:
forming a weighted sum of the resultant current in the first conductive line and the resultant current in the second conductive line; and
determining, from the weighted sum, a product of the input value for the neural network and the weight value of the neural network.
14. The method of claim 13 , wherein forming the weighted sum comprises:
weighting the resultant current in the first conductive line and the resultant current in the second conductive line according to a relative significance of the first bit of the multi-bit weight value and the second bit of the multi-bit weight value.
15. An apparatus, comprising:
one or more control circuits configured to connect to a plurality of memory cells each connected between one of a plurality of word lines and one of one or more bit lines, the one or more control circuits configured to:
receive an N-bit input value for a neural network;
convert the input value into a plurality of voltage values;
concurrently apply each of the voltage values to a corresponding one of the word lines; and
determine a product of the N-bit input value for the neural network and a first bit value of a weight value for the neural network by measuring a resultant current in a first bit line connected to a first subset of the memory cells each storing the first bit value and each connected between a corresponding one of the plurality of word lines and the first of the bit line.
16. The apparatus of claim 15 , wherein the plurality of voltage values are N voltage values, each corresponding to one bit of the N-bit input value, and wherein the one or more control circuits are configured to concurrently apply each of the N voltage values to a corresponding one of the word lines.
17. The apparatus of claim 15 , wherein the one or more control circuits are further configured to:
concurrently with determining the product of the N-bit input value for the neural network and the first bit value of the weight value for the neural network, determine a product of the N-bit input value for the neural network and a second bit value of the weight value for the neural network by measuring a resultant current in a second bit line connected to a second subset of the memory cells each storing the second bit value and each connected between a corresponding one of the plurality of word lines and the second of the bit line.
18. The apparatus of claim 17 , wherein the one or more control circuits are further configured to:
form a weighted sum of the resultant current in the first bit line and the resultant current in the second bit line; and
determine, from the weighted sum, the product of the input value for the neural network and the weight value of the neural network.
19. The apparatus of claim 18 , wherein, in forming the weighted sum, the one or more control circuits are further configured to:
weight the resultant current in the first bit line and the resultant current in the second bit line according to a relative significance of the first bit of the weight value and the second bit of the weight value.
20. The apparatus of claim 15 , wherein, in converting the input value into the plurality of voltage values, the one or more control circuits are further configured to:
receive a supply voltage; and
multiply a value of the supply voltage by each of multiplier values to thereby generate the N voltage values each corresponding to a bit of the input value, the values of the N multiplier values corresponding to a relative significance of the corresponding bit.