# Edge AI without Compromise: Efficient, Versatile and Accurate Neurocomputing in Resistive Random-Access Memory

**Authors:** Weier Wan, Rajkumar Kubendran, Clemens Schaefer, S. Burc Eryilmaz, Wenqiang Zhang, Dabin Wu, Stephen Deiss, Priyanka Raina, He Qian, Bin Gao, Siddharth Joshi, Huaqiang Wu, H. -S. Philip Wong, Gert Cauwenberghs  
**arXiv:** [2108.07879](https://arxiv.org/abs/2108.07879)  
**Date:** 2021-08-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.07879v1/

## Key points

- Edge hardware development is crucial for cloud-level AI functionalities at the edge of the internet
- NeuRRAM is the first multimodal edge AI chip using RRAM CIM
- NeuRRAM delivers high versatility and energy-efficiency 5-8 times better than prior art across various computational bit-precisions
- Inference accuracy is comparable to software models with 4-bit weights on all measured standard AI benchmarks, including impressive results on MNIST image classification, CIFAR-10 image classification, and Google speech command recognition
- NeuRRAM achieved a 70% reduction in image reconstruction error on a Bayesian image recovery task
- These results pave the way towards building highly efficient and reconfigurable edge AI hardware platforms for more demanding and heterogeneous AI applications in the future.

## Summary

The development of edge hardware capable of processing multiple modalities of sensory data at unprecedented energy-efficiency is crucial for realizing cloud-level artificial intelligence functionalities directly on distributed devices at the edge of the internet. To address trade-offs between efficiency, versatility, and accuracy, researchers have co-optimized across all hierarchies of design from algorithms and architecture to circuits and devices to create NeuRRAM - the first multimodal edge AI chip using resistive random-access memory (RRAM) CIM. This innovative chip delivers a high degree of versatility for diverse model architectures while recording energy-efficiency 5-8 times better than prior art across various computational bit-precisions. Inference accuracy is comparable to software models with 4-bit weights on all measured standard AI benchmarks including an impressive 99.0% accuracy on MNIST image classification, 85.7% accuracy on CIFAR-10 image classification, and 84.7% accuracy on Google speech command recognition. Additionally, NeuRRAM achieved a 70% reduction in image reconstruction error on a Bayesian image recovery task. These results pave the way towards building highly efficient and reconfigurable edge AI hardware platforms for more demanding and heterogeneous AI applications in the future.