Close

Presentation

Evolved Neural-Hardware-Compiler Co-Design
TimeThursday, December 9th11:00am - 11:30am PST
Location3014
Event Type
Research Manuscript
Virtual Programs
Presented In-Person
Keywords
AI/ML System Design
Topics
Design
DescriptionHardware architecture search is an efficient tool to enable specialized accelerator designs for neural networks. Current search frameworks focus on sizing the architectural hyper-parameters while neglect searching the PE connectivity and compiler mappings.
To tackle this challenge, we propose ENHCS that holistically searches the neural/hardware architectures and compiler mapping in one optimization loop. ENHCS composes highly matched architectures together with efficient mapping for given neural networks. With the same computation resources, EHSS can rival the human design Eyeriss by 4.4X EDP reduction with 2.7% accuracy improvement on ImageNet and offers 3.5X EDP reduction than only searching the architectural sizing.