Neural Architecture Search (NAS)-Based AI Accelerator Design with Memristor Crossbars
Grant Data
Project Title
Neural Architecture Search (NAS)-Based AI Accelerator Design with Memristor Crossbars
Principal Investigator
Professor Wong, Ngai
(Principal Investigator (PI))
Co-Investigator(s)
Professor Li Can
(Co-Investigator)
Professor Wang Zhongrui
(Co-Investigator)
Professor Yu Hao
(Co-Investigator)
Duration
28
Start Date
2021-09-01
Amount
2367000
Conference Title
Neural Architecture Search (NAS)-Based AI Accelerator Design with Memristor Crossbars
Keywords
AI Accelerator Design, Memristor Crossbars, Neural Architecture Search
Discipline
Bioelectronics
Panel
Engineering (E)
HKU Project Code
MHP/066/20
Grant Type
Mainland-Hong Kong Joint Funding Scheme (MHKJFS)
Funding Year
2021
Status
On-going
Objectives
The use of analog resistive memory crossbars in implementing various neural networks has proven to be a highly promising R&D direction for low-power AI accelerator design. The processing-in-memory (PIM) architecture could mitigate the von Neuman bottleneck, showing its promise to be the workhorse for next-generation machine learning hardware. However, so far, the resistive random-access memory (RRAM) PIM system designs are still primitive rather than being task- and efficiency-oriented. To expedite the commercialization of this cutting-edge computing platform, this project proposes, for the first time, an algorithm-hardware co-design flow based on modified neural architecture search (NAS) to maximize the network performance and energy efficiency. With this customized flow, we target >30% energy efficiency than a non-NAS-optimized design and fabricate a >50 TOPS/W RRAM AI accelerator.