Efficient neural network accelerator dataflows
A distributed deep neural net (DNN) utilizing a distributed, tile-based architecture includes multiple chips, each with a central processing element, a global memory buffer, and a plurality of additional processing elements. Each additional processing element includes a weight buffer, an activation buffer, and vector multiply-accumulate units to combine, in parallel, the weight values and the activation values using stationary data flows.
1. A semiconductor neural network accelerator comprising a plurality of chips, each chip comprising:
a controller;
a global memory buffer;
a plurality of processing elements, each comprising a weight buffer, an accumulation memory buffer, and a plurality of vector multiply-accumulate units;
at least one of the vector multiply-accumulate units comprising:
a first collector interposed to receive weights from the weight buffer and to provide the weights for a convolution computation; and
a second collector interposed to receive results of the convolution computation and to provide the results to the accumulation memory buffer; and
configuration logic to configure a depth of one or both of the first collector and the second collector to implement a stationary data flow by the neural network during the convolution computation.
2. The semiconductor neural network accelerator of claim 1 , further comprising:
the configuration logic to configure a depth of one or both of the first collector and the second collector to implement multi-level weight stationary/output stationary data flows for the convolution computation.
3. The semiconductor neural network accelerator of claim 1 , further comprising:
the configuration logic to configure a depth of one or both of the first collector and the second collector to implement one or more of multi-level weight stationary, output stationary, and input stationary data flows for the convolution.
4. The semiconductor neural network accelerator of claim 1 , wherein the configuration logic configures at least two or more of the processing elements to:
compute a portion of the convolution and forward the portion to neighboring processing elements for completion; and
communicate results of completion of the convolution to the global memory buffer.
5. The semiconductor neural network accelerator of claim 4 , configured such that results of completion of the convolution are staged in the global memory buffer for communication between layers of the neural network.
6. The semiconductor neural network accelerator of claim 1 , one or more of the processing elements further comprising:
a third collector disposed between an activation buffer and the vector multiply-accumulate units.
7. The semiconductor neural network accelerator of claim 6 , further comprising:
the configuration logic to configure a depth of the third collector to adjust a stationary level of input activations during the convolution computation by the vector multiply-accumulate units.