Other and custom chips
SambaNova, Maia, MTIA, Ascend and other accelerators. 44 providers on EveryNeocloud so far.
Beyond Nvidia, AMD and the chips with their own pages here, many companies and countries run their own accelerators. SambaNova sells its RDU; Microsoft designs Maia, Meta designs MTIA, and Tesla built the D1 for Dojo. In China, Huawei's Ascend, Baidu's Kunlun and Alibaba's chips matter more each year under US export controls. Japan's Fugaku runs on Fujitsu's Arm processors.
Why build your own
A company that buys hundreds of thousands of GPUs can save money and gain control by designing chips for its own models. Custom chips can drop features a general GPU must carry, but they need years of work, a supplier such as TSMC or SMIC to make them, and software that developers will use. Broadcom and Marvell help several firms design them.
The Chinese case
US export rules since 2022 have limited China's access to the best Nvidia and AMD chips. Chinese clouds have turned to Huawei's Ascend 910 series, Baidu's Kunlun, Alibaba's T-Head designs and others, made at SMIC with older processes. Earlier, the Sunway TaihuLight supercomputer of 2016 used Chinese processors throughout. Supporters of the rules say they slow China's military AI; critics say they speed up its own chip industry.
Famous other and custom chips providers
Still to be written: Sunway TaihuLight.
Further reading
Photo credits on this page
The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.
- Tesla Dojo: Steve Jurvetson, CC BY 2.0, source
- Amazon Web Services: Raysonho @ Open Grid Scheduler / Grid Engine, CC0, source
- Alibaba Cloud: Windmemories, CC BY-SA 4.0, source
- Baidu: N509FZ, CC BY-SA 4.0, source
- Amazon Elastic Compute Cloud: Raysonho @ Open Grid Scheduler / Grid Engine, CC BY-SA 3.0, source
- Fugaku: 文部科学省, CC BY 4.0, source
- Google data centers: erwinboogert, CC BY-SA 3.0, source
