Presentation
Specialization meets Flexibility: a Heterogeneous Architecture for High-Efficiency, High-Flexibility AR/VR Processing
DescriptionEmerging AR-VR applications require complex heterogeneous workloads, mixing Deep-Learning and DSP tasks, on devices embedded in the frame of eyeglasses. We propose ArchiMEDES, a heterogeneous-SoC platform with a programmable cluster of RISC-V cores alongside a configurable DNN engine (NEureka). ArchiMEDES features a low-overhead Heterogeneous Cluster Interconnect to enable fast RISC-V/NEureka cooperation. We show post-layout results targeting 22nm technology;
ArchiMEDES shows a peak combined performance of up to 1.14 TMAC/s at an efficiency of 4.38 TMAC/s/W. Hardware-Software cooperation in ArchiMEDES enables an 2× speedup and 1.5×
efficiency boost in an AR-VR gaze tracking case study, compared to a non-heterogeneous RISC-V cluster
ArchiMEDES shows a peak combined performance of up to 1.14 TMAC/s at an efficiency of 4.38 TMAC/s/W. Hardware-Software cooperation in ArchiMEDES enables an 2× speedup and 1.5×
efficiency boost in an AR-VR gaze tracking case study, compared to a non-heterogeneous RISC-V cluster
Event Type
Research Manuscript
TimeTuesday, July 11th10:55am - 11:10am PDT
Location3006, 3rd Floor
Design
RISC-V
SoC, Heterogeneous, and Reconfigurable Architectures