Presentation
Map-and-Conquer: Energy-Efficient Mapping of Dynamic Neural Nets onto Heterogeneous MPSoCs
DescriptionHeterogenous SoCs comprise diverse processing units of varying compute capabilities. To date, neural networks (NNs) mapping schemes onto such systems are yet to exploit the full potential of processing parallelism, made possible through both the intrinsic NNs structure and the underlying hardware composition. In this paper, we present a novel methodology to achieve that, where contrary to conventional mapping approaches that partition NNs according to their `depth', we propose to partition them according to the `width' dimension as concurrent mappable workloads. Our experiments using a variety of hardware and NN architectures have shown significant performance gains can be achieved.
Event Type
Research Manuscript
TimeTuesday, July 11th10:40am - 10:55am PDT
Location3006, 3rd Floor
Design
RISC-V
SoC, Heterogeneous, and Reconfigurable Architectures