Recently, representative array-level demonstrations have been reported. These include IBM’s 500 × 661 phase change memory array for handwritten-digit recognition using the Modified National Institute of Standards and Technology (MNIST) database (52), Tsinghua’s 128 × 8 analogue resistive RAM array for face recognition (53), UCSB’s 12 × 12 crossbar array for pattern recognition (54), and UCSB’s floating-gate array for MNIST image recognition (55)
The energy efficiency of the system was over 119.7 trillion equivalent operations per second per watt using
a readout of 10ns, and this is expected to increase significantly with larger vectors and matrices and with improvements in circuitry
## introductory video - [The AI Hardware Problem](https://www.youtube.com/watch?v=owe9cPEdm7k)
[Mapping the 100bn IT spending opportunity](https://knowen-production.s3.amazonaws.com/uploads/attachment/file/3076/AI%2Bhardware.pdf)
2018 [Big Bets on A.I. Open a New Frontier for Chip Start-Ups, Too](https://knowen-production.s3.amazonaws.com/uploads/attachment/file/5036/chip%2Bstartups.pdf)
advanced discussion [Future of AI Hardware Panel Dave Patterson, Bryan Catanzaro, Andrew Feldman, & Cade Metz](https://www.youtube.com/watch?v=E__85F_vnmU)
## technical [Does AI have a hardware problem?](https://www.nature.com/articles/s41928-018-0068-2)
> As deep neural networks continue to improve and grow, innovations in hardware will be required in order to meet the increasing computational demands.
## 2018 +++ [Scaling for edge inference of deep neural networks](https://knowen-production.s3.amazonaws.com/uploads/attachment/file/5038/scaling%2Bfor%2Bedge%2Binference.pdf) Xiaowei Xu
- page 218, see references 50-56
> Recently, representative array-level demonstrations have been reported. These include IBM’s 500 × 661 phase change memory array for handwritten-digit recognition using the Modified National Institute of Standards and Technology (MNIST) database (52), Tsinghua’s 128 × 8 analogue resistive RAM array for face recognition (53), UCSB’s 12 × 12 crossbar array for pattern recognition (54), and UCSB’s floating-gate array for MNIST image recognition (55)
## 2017 +++ Can Li [Analogue signal and image processing with large memristor crossbars](https://knowen-production.s3.amazonaws.com/uploads/attachment/file/5184/li2017.pdf) **R. Stanley Williams**
> The energy efficiency of the system was over 119.7 trillion equivalent operations per second per watt using
> a readout of 10ns, and this is expected to increase significantly with larger vectors and matrices and with improvements in circuitry
## the market
Podcast [The Future of AI Hardware with James Wang](https://podcasts.apple.com/gb/podcast/fyi-for-your-innovation/id1271691895?i=1000443083259) FYI - For Your Innovation
- NVIDIA
2020 [Beyond NVIDIA — The Fast-Changing Specialized AI Compute Hardware Market](https://blog-idcuk.com/the-fast-changing-specialized-ai-compute-hardware-market/)
- Graphcore
https://www.graphcore.ai/posts/ai-research-directions-in-2021
## very GENERAL - https://a16z.com/2019/11/15/the-end-of-cloud-computing-2/
https://www.technologyreview.com/2021/02/24/1014369/10-breakthrough-technologies-2021/
https://petewarden.com/2018/06/11/why-the-future-of-machine-learning-is-tiny/
2020 [The Edge Computing Opportunity: It’s Not What You Think](https://blog.cloudflare.com/cloudflare-workers-serverless-week/) Matthew Prince
