Better mass produce before TSMC gets ban notice.
Chinese semiconductor industry
- Thread starter Hendrik_2000
- Start date
- Status
Better mass produce before TSMC gets ban notice.
It's 寒武纪科技 in Chinese. Literally "Cambrian technology." The company is a spin off from CAS AI chip programme, in which the initial AI accelerator chip released was called "Cambrian." They named it after the Cambrian explosion, when occurance of diverse multicellular life exploded on Earth. CAS hoped that its project would spur widespread Chinese AI chip development, which it did. No small thanks to the comrades across the Pacific though.
Last edited:
US can't ban all firms. TSMC , Samsung and all semiconductor manufacturing tool makers across the globe heavily dependent on China's import. coz PRC is the largest importer of chips worth over $300 billion annually. everything will collapse if they try to cut China off. this is why US doing targeted sanctions.
Hope for the best and prepare for the worst.
process should be, get as many large players signed as possible -> have TSMC produce start mass producing -> deliver BR100/104 to customers for testing -> sort out bugs (believe me, there will be a lot) -> get larger firm orders/contract and sign up as many large players as possible -> place a huge order with TSMC to pre-empt sanctions -> work with SMIC 7 nm+ process (said to be comparable to Samsung 5nm) to develop next generation GPU -> develop more AI, self driving and server chips.
btw, did a little more digging around. Looks like Inspur is going full board with domestic players. Here they signed up contract with Kunlun-2 back in July. Maybe they heard from their sources that US might ban AI chips, so signed up Kunlun-2 and BR-100 in back to back month. To me, Kunlun-2 and BR-100 are complementary, so it makes total sense to source from both. Looks like Baidu is sourcing both also.
Didn't realize this, but Inspur has over 50% market share in China's AI server market. No wonder they got blacklisted by DoD.
other quotes I've seen online.
@european_guy mentioned earlier that Nvidia sells A100 in 8x boxes. Looks like what Biren technology built here with Inspur is similar to what Inspur built with Nvidia A100. this should be ready for sampling in Q4.
If you read their website, this hearten server with 8 BR100 actually achieves 8x the computation of each BR100. I mean if Inspur can help Biren develop and test/trial out this hearten server, it's going to be quite the beast. It claims to be 2.4 to 3x the processing power of A100. Who needs A100 or even H100 when you have this.
Looks like their software platform SUPA should be similar to Nvidia's CUDA. That would be useful in encouraging people to switch over. Knowing how some of my colleagues have spent hours/hours debugging/writing to new FPGAs, I would imagine most AI developers are not keen on trialing new hardware and new APIs (so similarities do help). Once you make the commitment to a platform, it's pretty sticky. Hard to move away.
PyTorch - developed by Meta AI.
TensorFlow - developed by Google.
PaddlePaddle - developed by Baidu
Back to Kunlun-2, some of their likely partners/customers aside from Inspur
This is Nanjing future web.
There is UDG China
Also Apeman walk. (I cannot believe they call themselves that)
Also chongqing high court, lol. I guess they need AI chips too
So definitely seems like Biren is capturing the higher end market whereas Kunlun-2 at this point is aimed at lower end than that.
Is it open sourced? If not then it better be, because ts a pain in the ass to have to learn how to use proprietary software for each AI hardware out there.
Either open source it and make an "alliance" with other companies to use the same software platform or its going to be a mess. 100 companies -> 100 APIs
They should at the bare minimum design a domestic open standard like Vulkan or opengl does for graphics and certain gpu compute.
Thanks for doing all that reading and summarizing for us.
What else can we say?
The future is now. Tech decoupling is happening now, and China will do that seamlessly as possible.
Inspur is a big name. That is instant creditably. Aren't they the best company in the world building server racks? If the best server maker in the world wants your chips, then everyone will want your chips.
Also, seems kind of clear what they are doing.
This is an arms race in technology, more specifically in platform technology and AI technology.
All these platforms, is PaaS, SaaS, IaaS, they need servers, and the better the servers will make these PaaS, SaaS, IaaS, better too.
Then layer on top of that the AI, then make the systems operations better still, in terms of efficiency and speed and potential I guess.
(Note, that much of this arms race in technology, is not even against the Americans. The Chinese company is trying to improve themselves in their battles with other Chinese companies.)
So if this is an arms race in platform technology, where now they are going to implement AI to make those platforms better, then we better need the best servers we can find, using chips such as NIvaida chips, Biren chips.
If your platform cannot compete, then you could be toast.
(TikTok itself could be their own platform, but that is another topic for another day.)
Chinese tech is truly cutting edge. Horizontally and vertically integrated.
All that remains is the leading edge fab to be done. All of the other major things needed, seem to be there.
Wow!
Wow!
That is why these recent bans from the US government, about EDA and the GPU, seems so weird and out of place.
If this was cut off tomorrow, nothing will stop in China.
Those platforms are still up and running. Those servers are still crunching data.
If the idea so to prevent the Chinese from reaching the future, it too apparent the Chinese will reach the future with or without American inputs.
The tech war has gone coo-coo.
- Status