Authors: Minki Jeong, Wanyeong Jung

Published on: July 16, 2022

Impact Score: 8.15

Arxiv code: Arxiv:2207.07862

Summary

• What is new: MAC-DO, a novel DRAM-based in-situ accelerator that supports multi-bit MAC operations within a single cycle using a new analog computation method.
• Why this is important: Existing DRAM-based in-situ accelerators face low array utilization and inefficiency in performing complex operations like multi-bit MAC, leading to poor power efficiency and high data movement.
• What the research proposes: The paper introduces a MAC-DO cell composed of two 1T1C DRAM cells, which enables efficient, concurrent multi-bit MAC operations within a single cycle, significantly improving throughput and energy efficiency.
• Results: A MAC-DO array can accelerate matrix multiplications efficiently, supporting computations in deep neural networks with minimal data movement and high throughput.

Technical Details

Technological frameworks used: nan

Models used: MAC-DO uses charge steering for analog computation and achieves compatibility with existing 1T1C DRAM cells.

Data used: nan

Potential Impact

This technology could disrupt markets relying on deep learning and artificial intelligence, benefiting semiconductor companies and various tech industries implementing DNNs.

Want to implement this idea in a business?

We have generated a startup concept here: ChipsAI.