Resource Efficient Hardware Implementation of Multi-Variate Regression Training

Multi-variate linear regression is one of the most popular supervised machine learning algorithms that is used to find a linear relation between independent variables and a single dependent variable. It has been successfully utilized in predicting dependent variables in many applications. Due to the wide adoption of machine learning algorithms, researchers have been developing ever more efficient implementations of such algorithms. Hardware implementations can offer high computational power and parallel processing. Field Programmable Gate Arrays (FPGAs) offer a very attractive hardware implementation option. FPGAs offer a flexible architecture that allows for algorithm modifications as needed. They can outperform GPUs and CPUs in terms of speed and power efficiency. In this paper, we propose a resource-efficient hardware implementation of a multi-variate linear regression training circuit on FPGA using Verilog HDL. This design leverages the available Block RAMs (BRAMs) on the FPGA board to store the data samples before feeding them to the algorithm. It is parametrizable in terms of both the number of data samples and the number of features in the data. However, a re-synthesis of the design is required when the number of features changes. We tested the design using a 200-training data set with 10 features and the time needed to produce the coefficients was 2.101 milliseconds. We synthesized the design using a low-cost FPGA board with a Xilinx xc7a100t-3-fgg484 FPGA device. The design only utilized 4% of the available registers, 8% of the available Logic Units, 5% of the available block RAMs, and 10% of the available DSP48E1 blocks. The estimated average power used was 130.84 milliwatts. A Python implementation of the same algorithm took 11.5921 milliseconds to compute the results using an Intel Core i7-10510U Processor with an average power of 15 W. Hence, the hardware implementation achieved a 5.5x reduction in computation time and a 115x reduction in power consumption.