Presentation
ViTSen: Enabling Vision Transformers at the Edge Through In/Near -Sensor Processing Schemes
DescriptionThis paper introduces ViTSen, optimizing Vision Transformers (ViTs) for resource-constrained edge devices. It features an in-sensor image compression technique to effectively reduce data conversion and transmission power costs. Further, ViTSen incorporates a ReRAM crossbar array, enabling efficient near-sensor analog convolution. This integration and novel pixel reading and peripheral circuitry decrease the reliance on analog buffers and converters, significantly lowering power consumption. To make ViTSen compatible, several established ViT algorithms have undergone quantization and channel reduction. Circuit-to-application co-simulation results show that ViTSen maintains accuracy comparable to a full-precision baseline across various data precisions, achieving an efficiency of approximately ~3.1 TOp/s/W.
Event Type
Work-in-Progress Poster
TimeWednesday, June 265:00pm - 6:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
IP
Security