Deep learning chips use difference math to boost speed and save energy
PENDA: An Efficient Processing Element via Norm-of-Difference for Deep Learning Accelerators
Hardware Architecture
Summary
Calculations called inner products take up a lot of time and power in deep learning, which is how computers learn from data. Most ways to speed this up guess the answers a little, which can cause mistakes. The authors came up with a new way using a math rule called the law of cosines to change multiplication into a subtraction-based operation. This new method saves chip space, energy, and makes the chip work faster without losing accuracy. It helps make deep learning hardware more efficient and reliable.
deep learninginner productmultiplicationlaw of cosineshardware acceleratorprocessing elementenergy efficiencyclock periodsquared differencemultiply-accumulate unit
Authors
Kai-Chieh Hsu, Tian-Sheuan Chang
Abstract
Inner product computation dominates the computational cost of deep learning models; thus, accelerating this primitive is key to improving hardware efficiency. However, most existing techniques rely on approximations, which can degrade model accuracy. To preserve exactness while optimizing hardware, this paper presents PENDA (processing element via norm-of-difference architecture), which leverages the law of cosines to recast multiplications as squared-difference operations. Replacing multiply-accumulate units with the proposed norm-of-difference units yields 11~36%, 5~48%, and 11~19% reductions in area, energy, and clock period, respectively, for the PE array of a deep learning accelerator.