In this paper submitted by Jeffrey Ullman, the focus is on matrix multiplication and its impact on machine-learning algorithms. The author highlights the need for accelerator chips that can perform matrix multiplication faster than traditional processors or even GPUs. The paper presents a unique approach to matrix multiplication, eliminating the need for a scalar multiplier circuit. By using only a single addition and on-chip copy operation, the proposed method offers a more efficient and energy-saving alternative. With this approach, it becomes possible to design a matrix-multiplier chip that can accommodate more processors, leading to a significant speedup.
https://arxiv.org/abs/2307.01415