COSAIM: Counter-based Stochastic-behaving Approximate Integer Multiplier for Deep Neural Networks
TimeTuesday, December 7th4:10pm - 4:30pm PST
Event Type
Research Manuscript
Virtual Programs
Presented In-Person
Approximate Computing for AI/ML
DescriptionWe propose a new stochastic-behaving approximate integer unsigned multiplier, COSAIM. The new design is an improved stochastic multiplier. Our evaluation shows that the COSAIM with error improvement can achieve error bias, mean error, peak errors at 0.06%, 0.3%, 1.81%. Experimental results show that COSAIM can save up to 53.95%, 32.84%, 52.24%, 21.05% in area, power, energy and the product Area•1/Throughput, respectively. By doing parallel design, COSAIM can further lead to improvements in area, power and energy reduction by 60.44%, 53.33% and 68.54%, respectively. When implemented in CNN application, COSAIM can achieve similar inference accuracy compared to the baseline.