Does dataflow matter for DNN accelerator performance?
A huge amount has been published about deep neural network accelerator designs, but what parts of the design actually matter?
A huge amount has been published about deep neural network accelerator designs, but what parts of the design actually matter?