Multi-memory on-chip computational network
Provided are systems, methods, and integrated circuits for neural network processing. In various implementations, an integrated circuit for neural network processing can include a plurality of memory banks storing weight values for a neural network. The memory banks can be on the same chip as an array of processing engines. Upon receiving input data, the circuit can be configured to use the set of weight values to perform a task defined for the neural network. Performing the task can include reading weight values from the memory banks, inputting the weight values into the array of processing engines, and computing a result using the array of processing engines, where the result corresponds to an outcome of performing the task.
1. An integrated circuit comprising:
a first array of processing engines; and
a plurality of memory banks for storing a set of weight values for a neural network, wherein each memory bank from the plurality of memory banks is independently accessible, wherein the plurality of memory banks and the first array of processing engines are on a same die, and wherein the plurality of memory banks supports simultaneous accesses including reading a first value from a first memory bank, while a second value is being read from or written to a second memory bank;
wherein, upon receiving input data, the integrated circuit is configured to use the set of weight values to perform a task defined for the neural network, and wherein performing the task includes:
reading weight values from the plurality of memory banks;
inputting the weight values and the input data into the first array of processing engines; and
computing a result using the first array of processing engines, wherein the result corresponds to an outcome of performing the task.
2. The integrated circuit of claim 1 , wherein each of the first and second values includes one of a weight value, an input value, or an intermediate result.
3. The integrated circuit of claim 1 , further comprising:
a second array of processing engines, wherein a first set of memory banks from the plurality of memory banks is initially configured for use by the first array of processing engines, wherein a second set of memory banks from the plurality of memory banks is initially configured for use by the second array of processing engines, and wherein the first set of memory banks and the second set of memory banks each include a portion of the set of weight values.
4. The integrated circuit of claim 3 , wherein performing the task further includes:
computing, by the first array of processing engines, an intermediate result, wherein the first array of processing engines computes the intermediate result using weight values from the first set of memory banks; and
reading, by the first array of processing engines, additional weight values from the second set of memory banks, wherein the first array of processing engines uses the intermediate result and the additional weight values to compute the result.
5. The integrated circuit of claim 3 , wherein the set of weight values occupies less than all of the second set of memory banks, and wherein the second array of processing engines performs computations using a part of the second set of memory banks that is not occupied by the set of weight values.
6. The integrated circuit of claim 1 , wherein performing the task further includes:
determining that an amount of memory needed for storing an intermediate result has decreased;
reading an additional set of weight values from another memory; and
storing the additional set of weight values in a portion of the plurality of memory banks that is no longer needed for the intermediate result.
7. The integrated circuit of claim 1 , wherein the plurality of memory banks includes at least as many memory banks as a number of rows in the first array of processing engines.
8. The integrated circuit of claim 1 , wherein the plurality of memory banks includes at least as many memory banks as a number of columns in the first array of processing engines.
9. The integrated circuit of claim 1 , further comprising a pooling circuit to perform a pooling function on an output of the first array of processing engines to generate an intermediate result before writing the intermediate result to a portion of the plurality of memory banks.
10. The integrated circuit of claim 1 , further comprising an activation circuit to perform an activation function on an output of the first array of processing engines to generate an intermediate result before writing the intermediate result to a portion of the plurality of memory banks.
11. A method comprising:
storing a set of weight values in a plurality of memory banks of a neural network processing circuit, wherein the neural network processing circuit includes an array of processing engines on a same die as the plurality of memory banks, wherein the plurality of memory banks supports simultaneous accesses including reading a first value from a first memory bank, while a second value is being read from or written to a second memory bank, and wherein the set of weight values is stored prior to receiving input data;
receiving input data; and
using the set of weight values to perform a task defined for a neural network, wherein performing the task includes:
reading weight values from the plurality of memory banks;
inputting the weight values and the input data into the array of processing engines; and
computing a result using the array of processing engines, wherein the result corresponds to an outcome of performing the task.
12. The method of claim 11 , wherein each of the first and second values includes one of a weight value, an input value, or an intermediate result.
13. The method of claim 11 , wherein performing the task further includes:
computing, by the array of processing engines, an intermediate result, wherein the array of processing engines computes the intermediate result using weight values from a first set of memory banks in the plurality of memory banks.
14. The method of claim 13 , wherein performing the task further includes:
writing the intermediate result to a first portion of the plurality of memory banks, and moving the intermediate result to a second portion of the plurality of memory banks from which the intermediate result is read.
15. The method of claim 13 , wherein performing the task further includes:
reading, by the array of processing engines, additional weight values from a second set of memory banks in the plurality of memory banks, wherein the array of processing engines uses the intermediate result and the additional weight values to compute the result.
16. The method of claim 15 , wherein a second array of processing engines performs computations using a part of the second set of memory banks that is not occupied by the set of weight values.
17. The method of claim 11 , wherein performing the task further includes:
determining that an amount of memory needed for storing an intermediate result has decreased;
reading an additional set of weight values from another memory; and
storing the additional set of weight values in a portion of the plurality of memory banks that is no longer needed for the intermediate result.
18. The method of claim 11 , wherein each row of the array of processing engines reads from a respective memory bank of the plurality of memory banks.
19. The method of claim 11 , wherein each column of the array of processing engines writes to a respective memory bank of the plurality of memory banks.
20. The method of claim 11 , further performing an activation function or a pooling function on an output of the array of processing engines to generate an intermediate result before writing the result to a portion of the plurality of memory banks.