Presentation
Additive Partial Sum Quantization
DescriptionDNN accelerators, significantly advanced by model compression and specialized dataflow techniques, have marked considerable progress. However, the frequent access of high-precision partial sums (PSUMs) leads to excessive memory demands in architectures utilizing weight/input stationary dataflows. Traditional compression strategies have typically overlooked PSUM quantization, a gap recently explored in compute-in-memory research. Moreover, these approaches are mainly toward reducing the Analog-to-Digital Converter (ADC) overhead, neglecting the critical issue of intensive memory access. This study introduces a novel Additive Partial Sum Quantization (APSQ) method, seamlessly integrating PSUM accumulation into the quantization framework. We further propose a grouping strategy that combines APSQ with PSQ enhanced by a floating-point regularization technique to boost accuracy. The experiments indicate that APSQ can efficiently compress PSUMs to INT-8, incurring a negligible degradation of accuracy for the Segformer-B0 and EfficientViT-B0 on the challenging Cityscapes dataset. This leads to a notable reduction in energy costs by 30~45%.
Event Type
Work-in-Progress Poster
TimeTuesday, June 256:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
IP
Security