Interesting thing from NIO earnings call today about the Nvidia ban.
The
will have no impact on NIO's operations in the short term, and in terms of AI computing power, NIO's current resources are sufficient for self-driving algorithm training.
There are local companies in China with similar chips, and NIO will evaluate their solutions.
There you have it, NIO and other companies can still do their computation with domestic chips.
I took a look at what they are talking about. It's well known that NIO had been working with Baidu cloud for a data center solution using A100 chips. I've also posted a lot about Baidu's work with Kunlun-2 and their participation in BR100 project.
A couple of things from this. Baidu builds a lot of AI cloud for big customers. It can make the sales pitch in China as the only service provider that can offer self developed AI IaaS layer including the Kunlun chip, Paddle Paddle deep learning platform and wnexin big data model.
A lot of this was unveiled as part of their new "cloud intelligence 3.0" platform. here is the key
the chip has been able to replace foreign-made chips and reduce costs by as much as 65%. At present, Kunlun (3rd generation) is already under R&D and it is expected to be mass-produced in 2024. The chip will become a substitute product for domestic high-end demand.
Nvidia is in huge trouble if Chinese domestic chips can reduce cost by 65%! I assume that Kunlun-3 will be higher end like BR100. Although, even the Kunlun-2 offers higher computation and low power consumption than the previous generation V100 from Nvidia. Nvidia doesn't have a market in China for AI accelerators if all they can export is V100.
Again, Kunlun-3 to enter mass production in start of 2024. It is expected to be able to replace high end GPUs in China. They also discussed the Beige 2.0 AI computation platform which uses Kunlun chip and improved AI training efficiency by 2 times and reduced self driving mass production from months to weeks.
So I took a look at what they talked about with integrating with other platforms. Here it is. I would assume NIO will probably opt for this with A100 no longer being available. This seems readily available and has good performance (although not as good as BR100).
This is a project Baidu cloud had with Jiangsu bank. It used 2/4/8 K200 (kunlun2) GPUs, FT2500-64 Feting CPU (or S2500?), PaddlePaddle ARM framework and PaddlePadle deep learning platform. It uses the Ubuntu Kylin OS. So, everything being used in this AI accelerator platform is made in China. Entirely sanction free.
This is what I can find on FT S2500. I'd be curious what they can make with the SMIC n+1 process since this came out in a year or 2 ago with what I assume to be the best SMIC could produce back then?
S2500 uses 16nm process, 64 cores, 2.1 GHz.
It seems like this is entirely developed to be sanction free.