Neural network accelerator with parameters resident on chip
One embodiment of an accelerator includes a computing unit; a first memory bank for storing input activations and a second memory bank for storing parameters used in performing computations, the second memory bank configured to store a sufficient amount of the neural network parameters on the computing unit to allow for latency below a specified level with throughput above a specified level. The computing unit includes at least one cell comprising at least one multiply accumulate (“MAC”) operator that receives parameters from the second memory bank and performs computations. The computing unit further includes a first traversal unit that provides a control signal to the first memory bank to cause an input activation to be provided to a data bus accessible by the MAC operator. The computing unit performs computations associated with at least one element of a data array, the one or more computations performed by the MAC operator.
1. An accelerator for accelerating tensor computations, comprising:
a memory configured for storing a plurality of weights of a neural network layer;
a tensor unit used to retrieve an input to the neural network layer from a memory location, the input being an element of a data array; and
an operator configured to perform one or more computations associated with the neural network layer, the one or more computations including a multiply operation of the input to the neural network layer and a corresponding weight stored in the memory; and
wherein the one or more computations are executed below a threshold latency and above a threshold throughput, and wherein the memory, the tensor unit, and the operator are located on a same die.
2. The accelerator of claim 1 , wherein the memory is configured to store more than 100,000 weights of one or more neural network layers.
3. The accelerator of claim 1 , wherein the memory is configured to store more than 1,000,000 weights of one or more neural network layers.
4. The accelerator of claim 1 , wherein the memory is configured to store more than 100,000,000 weights of one or more neural network layers.
5. The accelerator of claim 1 , wherein the memory comprises SRAM.
6. The accelerator of claim 1 , further comprising a data bus accessible by the operator.
7. The accelerator of claim 6 , wherein the tensor unit is configured to provide a control signal to cause an input to the neural network layer to be provided to the data bus accessible by the operator.
8. The accelerator of claim 7 , further comprising another memory in which the input to the neural network layer is stored.
9. The accelerator of claim 8 , wherein the input to the neural network layer comprises a plurality of inputs to the neural network layer, and the operator comprises a plurality of operators.
10. The accelerator of claim 9 , wherein the other memory is configured to provide the data bus one-by-one broadcast of the inputs to the neural network layer such that each of the plurality of operators receives a respective input of the plurality of inputs.
11. The accelerator of claim 1 , wherein the input to the neural network layer comprises an input activation for a node on the neural network layer.