Artificial neural network acceleration on FPGA using custom instruction

In this paper, we present the acceleration of a pre-trained feedforward artificial neural network executing on a NIOS II processor. Without the use of hardware acceleration, a feedforward artificial neural network spends much of its execution time on the calculation of the activation function between layers, in this case, the hyperbolic tangent function. A speedup of 4.36 was achieved via a custom instruction approximating the value of tanh(x) through the use of a range addressable lookup table.