What question did this study set out to answer?

To develop an energy-efficient, programmable accelerator architecture that overcomes hardware virtualization overheads during full neural-network inference executions.

February 13, 2021

1 A Programmable Neural-Network Inference Accelerator Based on Scalable In-Memory Computing

View Full Paper

Ask AI

Bookmark

Authors

HJHongyang JiaTsinghua University MOMurat OzatayPrinceton University YTYinqi TangPrinceton University

Discussion

Member takes

Implication

Hardware evaluation demonstrates scalable mixed-signal in-memory computing across diverse neural networks, suggesting reduced virtualization overheads via compiler co-design.

Key Points

To develop an energy-efficient, programmable accelerator architecture that overcomes hardware virtualization overheads during full neural-network inference executions.
Fabricated a 16nm programmable accelerator array combining mixed-signal in-memory computing, digital near-memory computing, and localized buffering and control.
Co-designed hardware execution units with a custom software graph compiler to execute flexible mapping algorithms across spatial and temporal domains.
Mitigated hardware virtualization overheads associated with state swapping, data buffering, and inter-core communication.
Maintained in-memory computing energy efficiency and high matrix-vector multiplication throughput across complete executions of diverse neural networks.

Cite This Study

Jia et al. (2021) studied this question.

synapsesocial.com/papers/6a125a94c031bb6829a6650d https://doi.org/10.1109/isscc42613.2021.9365788

View Full Paper

Ask AI

Bookmark

Implication

Hardware evaluation demonstrates scalable mixed-signal in-memory computing across diverse neural networks, suggesting reduced virtualization overheads via compiler co-design.

1 A Programmable Neural-Network Inference Accelerator Based on Scalable In-Memory Computing

Discussion

Member takes

Implication

Key Points

Cite This Study

1 A Programmable Neural-Network Inference Accelerator Based on Scalable In-Memory Computing

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider