Presentation
DyBit: Dynamic Bit-Precision Numbers for Efficient Quantized Neural Network Inference
DescriptionTo speedup inference of deep neural networks (DNNs), accelerators with low-precision data formats are actively researched. A prominent challenge is to port low-precision models to the edge with similar performance to high-precision networks. This work targets a non-uniform and efficient hardware-software co-design riding on a run-length encoded data format called DyBit (dynamic bit). DyBit can dynamically adjust the bit-level precision and range through separate bit-fields to represent the range and distribution of DNN weights/activations. To improve the quantization granularity, a flexible mixed-precision DyBit hardware is co-designed to versatilely adapt to various DNNs with superior accuracy-latency tradeoffs versus existing counterparts.
Event Type
Work-in-Progress Poster
TimeTuesday, July 11th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security