Compute-in-memory systems and methods
An integrated circuit device may include programmable logic circuitry on a first integrated circuit die and memory that includes compute-in-memory circuitry on a second die. The programmable logic circuitry may be programmed with a circuit design that operates on a first set of data. The compute-in-memory circuitry of the memory may perform an arithmetic operation using the first set of data from the programmable logic circuitry and a second set of data stored in the memory.
1. An integrated circuit device comprising:
a first die configurable to be programmed by a configuration bitstream comprising a circuit design for an artificial intelligence (AI) application that operates on input data to generate output data; and
a second die comprising memory to store computational data, wherein the memory comprises compute-in-memory circuitry configurable to perform an arithmetic operation using the output data and the computational data, and wherein the first die is configurable to broadcast data to the compute-in-memory circuitry.
2. The integrated circuit device of claim 1 , wherein the computational data comprises AI predictors, AI weights, or AI forecasts.
3. The integrated circuit device of claim 1 , wherein the compute-in-memory circuitry comprises a controller configurable to set a sequence of the arithmetic operation.
4. The integrated circuit device of claim 1 , wherein the first die is configurable to access the memory of the second die.
5. The integrated circuit device of claim 4 , wherein the memory of the second die comprises vertically stacked memory.
6. The integrated circuit device of claim 1 , wherein the first die comprises second memory to store the output data.
7. The integrated circuit device of claim 1 , wherein the arithmetic operation comprises a dot product operation, an accumulation operation, a matrix operation, a vector operation, a convolution operation, or a combination thereof.
8. The integrated circuit device of claim 1 , wherein the compute-in-memory circuitry is configurable to operate on the input data in a pattern based on a type of arithmetic computation.
9. The integrated circuit device of claim 1 , wherein the first die is stacked on top of the second die in a three-dimensional (3D) configuration.
10. A method, comprising:
implementing, via processing circuitry, a circuit design comprising an artificial intelligence (AI) application on a first die via a configuration bitstream; and
initializing, via the processing circuitry, compute-in-memory circuitry on a second die based on loading computational data associated with the circuit design;
transmitting, via the processing circuitry, data from the first die to the second die based on broadcasting the data to the compute-in-memory circuitry on the second die; and
causing, via the processing circuitry, an arithmetic operation to be performed by the compute-in-memory circuitry of the second die using the computational data and the data to generate output data.
11. The method of claim 10 , wherein initializing, via the processing circuitry, the compute-in-memory circuitry comprises:
retrieving the computational data from first memory of the first die or second memory of the second die.
12. The method of claim 11 , comprising transmitting, via the processing circuitry, the output data from the second die to the first die comprises gathering the output data from a plurality of compute-in-memory circuitry units of the compute-in-memory circuitry.
13. The method of claim 11 , wherein the arithmetic operation comprises a dot product operation, an accumulation operation, a matrix operation, a vector operation, a convolution operation, or a combination thereof.
14. The method of claim 10 , wherein:
the AI application comprises an artificial neural network function; and
the computational data associated with the circuit design comprises one or more neural network weights.
15. The method of claim 10 , wherein:
the circuit design comprises a filtering function; and
the computational data associated with the circuit design comprises one or more filter coefficients.
16. One or more tangible, non-transitory, machine-readable media comprising machine-readable instructions executable by one or more processors, wherein the instructions comprise instructions to:
receive a circuit design comprising an artificial intelligence (AI) application to be programmed onto a first die; and
generate a configuration bitstream to implement the circuit design comprising the AI application on the first die, wherein the configuration bitstream is to implement the AI application at least in part by:
implementing the circuit design comprising the AI application on the first die;
initializing compute-in-memory circuitry on a second die based on loading computational data associated with the circuit design;
transmitting data from the first die to the second die based on broadcasting the data from the first die to the compute-in-memory circuitry on the second die; and
causing an arithmetic operation to be performed by the compute-in-memory circuitry of the second die using the computational data and the data to generate output data.
17. The machine-readable media of claim 16 , wherein the configuration bitstream comprises an accelerator function configurable to perform machine learning, video processing, voice recognition, or image recognition.
18. The machine-readable media of claim 16 , wherein the instructions comprise instructions to:
cause the second die to transmit the output data to the first die.
19. The machine-readable media of claim 16 , wherein the instructions comprise instructions to:
generate a sequence of the arithmetic operation based on the circuit design; and
transmit an additional configuration bitstream comprising the sequence of the arithmetic operation to the second die.
20. The machine-readable media of claim 16 , wherein the compute-in-memory circuitry is configurable to operate as a dot product engine, and wherein the computational data comprises vectors and matrices.