Close

Presentation

Novel Adaptive Quantization Methodology for 8-bit Floating-Point DNN Training
DescriptionThere is a high energy cost associated with Deep Neural Networks training. Off-chip memory access contributes major portion to the overall energy consumption. Reduction in the number of off-chip memory-transactions can be achieved by quantization of the data words to low-bit-width (e.g., 8-bit). However, low-bit-width data formats suffer from a limited dynamic range. This paper presents a novel 8-bit Floating-Point (FP8) data format, which adapts to the desired range on-the-fly. Our approach is fully compatible with any DNN compute cores and quantizes only data to/from DRAM to FP8. the DRAM access energy is reduced by 3.07x with minor accuracy degradation.
Event Type
Work-in-Progress Poster
TimeWednesday, July 12th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
Topics
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security