Acceleration of tensor computations using the residual class system

The main scientific and practical barrier to the widespread dissemination of machine learning methods is the high computational complexity of tensor operations used in them. We propose a method for implementing tensor computations in a system of residual classes using table arithmetic for modular operations up to 8 bits wide, inclusive. Experimental modeling of the proposed method on FPGA Xilinx Spartan6 xc6slx9 showed that this method can be used to quickly organize computations when implementing tables on BRAM memory blocks. Modeling showed that the proposed approach allows us to accelerate the computations by a factor of two, compared with computations in the binary number system, which can be used to create hardware accelerators of tensor computations in practice.