An Efficient Dot-Product Unit Based on Online Arithmetic for Variable Precision Applications.

The dot-product operation is substantial to signal processing and AI algorithms; however, its efficiency is hampered by significant hardware resources and memory bandwidth required. To tackle these challenges, employing online arithmetic is a viable option. This bit-serial approach prioritizes the computation of the most significant digits of results, as opposed to conventional arithmetic, which begins with the least significant digit. Online arithmetic's suitability for variable precision computations allows for efficient termination of operations once desired precision or conditions are met, impacting both power consumption and performance positively. In spite of these merits, online arithmetic introduces an online delay, which is increased by the scale of the dot-product operation. To mitigate this, we present an online dot-product unit consolidating multiplications and additions within a single operation, ensuring a constant online delay regardless of the operation size. Leveraging shared partial sum registers in online multipliers reduces hardware costs. Moreover, our proposed design demonstrates a potential 2.03x speedup compared to state-of-the-art bit-serial dot-product units employed in deep neural network hardware accelerators, on average.