YOLoC: DeploY Large-Scale Neural Network by ROM-based Computing-in-Memory using ResiduaL Branch on a Chip
TimeThursday, July 14th2:37pm - 3pm PDT
Location3002, Level 3
Event Type
Research Manuscript
In-memory and Near-memory Computing
DescriptionComputing-in-Memory (CiM) is a promising technology to achieve higher energy efficiency in Matrix-Vector Multiplication. However, existing SRAM-CiMs, suffer from limited memory density, which will weaken the advantages of CiM. We proposed a ROM-based CiM that could perform MVM in a higher computable memory density array than SRAM. A fine-tune technology for ROM-CiM called Residual Compress is proposed to make ROM-CiM transfer to various tasks without retape-out. Evaluation in 28nm process shows 10x area saving and 30% energy efficiency improvement. Our proposed ROM-CiM has provided opportunities of enabling a new paradigm for future tiny-scale embedded object detection applications.