Presentation
ReARVR: A ReRAM-based DNN accelerator for mobile AR/VR devices
DescriptionAR/VR are the essential techniques for building the Metaverse.
However, AR/VR are typically closely coupled with the conventional computer graphics pipeline.
Prior DNN accelerators do not consider the complex interaction between the DNNs and the CGP in AR/VR.
In this work, we propose ReARVR to implements both the DNNs and the CGP on the same chip.
The benefits of the proposed ReARVR are threefold:
Firstly, computing the CGP in-situ is faster and more energy efficient than GPU.
Secondly, data movement is avoided between the boundaries of the DNNs and the CGP.
Finally, ReARVR promises faster execution for programmable CGPs.
However, AR/VR are typically closely coupled with the conventional computer graphics pipeline.
Prior DNN accelerators do not consider the complex interaction between the DNNs and the CGP in AR/VR.
In this work, we propose ReARVR to implements both the DNNs and the CGP on the same chip.
The benefits of the proposed ReARVR are threefold:
Firstly, computing the CGP in-situ is faster and more energy efficient than GPU.
Secondly, data movement is avoided between the boundaries of the DNNs and the CGP.
Finally, ReARVR promises faster execution for programmable CGPs.
Event Type
Work-in-Progress Poster
TimeWednesday, July 12th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security