BinDRAM: Binary neural network on unmodified commodity DRAM

Processing using memory leverages analog properties of memory infrastructure to implement logic and arithmetic operations. Processing in Commodity Off-The-Shelf (COTS) DRAM is especially promising due the massive parallelism across DRAM columns, and because it requires no modification to existing DRAM devices. One application that can benefit from such ubiquitous, easy to scale and affordable platform is artificial intelligence. Binary Neural Networks (BNNs) where weights and activations are constrained to binary values ( ± 1) are a highly efficient alternative to traditional neural networks. Such BNNs replace multiplication by XNOR which makes them a great target for using-memory implementation. Typical BNN, while binarizing linear layers (matrix-vector multiplication), still implements other network layers (batch normalization, softmax, and sometimes the input layer of convolutional neural network) in full precision. This limits the area and energy benefits, and requires architectural support for full precision operations. We propose BinDRAM, a true end-to-end binary network implemented in COTS DRAM that applies two ideas to eliminate full precision operations: (1) Duplicated last layer, where instead of a single input activation vector, a set of its slightly modified duplicates are processed by the output layer and their binary outputs are aggregated (for example, by simple majority) to render the final result; (2) implementing batch normalization by adding a constant to the POPCOUNT (i.e., the BNN dot product) and keeping the binary result. BinDRAM achieves significant throughput of 12,200 inferences/sec while consuming 0.89W. BinDRAM project is publicly available under https://anonymous.4open.science/r/BinDRAM.