Close

Presentation

XMAS: An Efficient Customizable Flow for Crossbarred-Memristor Architecture Search
TimeTuesday, July 12th5pm - 6pm PDT
LocationLevel 2 Exhibit Hall
Event Type
Engineering Track Poster
Engineering Tracks
Topics
AI
Back-End Design
Cloud
Embedded Systems
Front-End Design
IP
DescriptionMemristor Crossbar Arrays (MCAs) are customarily implemented in Processing-in-Memory (PIM) or In-Memory Computing (IMC) presently, which enables inference to be swift and power-efficient for specialized Neural Networks (NNs) and display the potential of memory wall assuagement. However, the co-optimization on Neural Architecture Search (NAS) of NNs and customizable hardware implementation demonstrate arduous search space and demanding hardware constraints. The search space traverses kernel size, depth, width, and input resolution. The hardware constraints extend to circuit topology and device variation, resulting in accuracy degeneration. To approach these challenges, we proposed a fast customizable method of crossbarred-memristor architecture search named XMAS. XMAS collectively delves into both the aforementioned search space with a special-crafted weighted sampling function and hardware constraints to scout the most robust and suitable NNs for customizable MCAs. Additionally, device variations are added in the search strategy rather than typically studied in evaluation. Different experiments are conducted on the CIFAR-10 dataset based on Resnet18 and Resnet50, which attain 85.2% and 88.3%.