VLSI@NU Researchers Wins Best Poster Award at IBM IEEE AI Compute Symposium

Northwestern researchers earned a best poster award at the 2020 IBM IEEE AI Compute Symposium, an annual conference that explores the latest developments in artificial intelligence ranging from hardware devices to software algorithms.

The best poster award went to “NCPU: A Binary Neural Network that Emulates RISC-V CPU at the Conjunction of Neuromorphic and Von-Neumann Architectures.” The work was presented by researchers from Northwestern’s VLSI@NU lab, including computer engineering PhD students Yuhao Ju and Tianyu Jia, and Northwestern Engineering's Jie Gu and Russell Joseph.

Fig. 1 A neural network-based CPU architecture where neurons are reconfigured to perform pipeline operations of CPU.

The work, demonstrated by a silicon test chip, proposes a novel and unique computing architecture that can be configured to a neural network accelerator for deep learning tasks or a central processing unit (CPU) microprocessor for conventional computing tasks. The combined architecture, referred to as “neural CPU,” uses a neural network to emulate conventional microprocessors, leading to 43 percent end-to-end performance gain or 74 percent energy savings for AI jobs on low-power edge computing devices.

The research offers a new computing architecture that combines the benefits of emerging neuromorphic computing and conventional Von-Neumann architecture. The detailed work was published at a recent IEEE/ACM International Symposium on Microarchitecture.

Fig. 2 Die photo of the fabricated chip in a 65nm CMOS technology and the testing setup for the chip.