Simple Neural Evaluation

The EE 272 Simple Neural Evaluation engine takes weights for a neural network, and applies them to input data. The design is flexible and assumes very little about the neural network. This design assumes all data is contiguous and all weights are processed. The design assumes the input data is 24 bits per sample in 2’s complement form. For math reasons, the input number are commonly scaled to be between -1 and +1. This implies the data is actually 8.24 binary fixed point. This scaling is important when processing weights. The weights are assumed to be in 8.24 form. (32 bits fixed binary point) The neural network can have up to 4096 inputs per layer, and this adds 12 additional bits to the sum. There can be up to 4096 nodes per layer. The input data is mapped to 8.24 before operation. The 8.24 x 8.24 multiplications result in 16.48 intermediate values which are summed to be 28.48 binary fixed point. After summing all products, the result is reduced to 8.24 by saturation. The data is then applied to a neural function. (a table lookup assumed to be the sigmoid or tanh function) To simplify the design, all neurons are assumed to be the same in each layer. The neuron result becomes the input to the next layer, or the output from the last layer. The design runs at a clock rate of 300 Mhz. (Use design ware multistage pipelined multipliers at this speed). The design performs 16 multiplies each cycle in the weight calculations. This requires a pipeline fetching weights and data, performing operations, and calculating the final sum. To support this performance level, data are stored in a memory which is 16*32 (512 bits wide). Each memory has up to 4M 512 bit words (24 bit address). The actual number of words is a system Verilog constant contained in the “[login to view URL]” file. There is one memory for calculation data, and one memory for weights and configuration data. The memory is pipelined and can only perform a read or write each clock, not both! Your design will require arbitration to decide if a read or write will happen to the memory.