China's DeepSeek Developing Its Own AI Chip (yahoo.com) 28
An anonymous reader quotes a report from Reuters: Chinese startup DeepSeek is developing its own AI chip, according to three people familiar with the matter, a push that could reduce its reliance on Nvidia and Huawei chips, which it has depended on to train and run its globally popular models. The chip is designed for inference -- the stage of AI computing in which a trained model generates responses for users -- rather than for training new models, the sources said. If successful, DeepSeek's expansion into semiconductor development would mark a major strategic shift for a company widely hailed in China as the country's AI champion, potentially adding to challenges faced by Chinese tech giant Huawei.
Re: (Score:2)
NVidia is an odd duck because they are profitable. And, they have now locked in long term contracts in business, industrial, automotive, aerospace and more. NVidia also has a surprisingly small footprint. Only 42000 employees. Facebook by random comparison has 77000, IBM has 264,000. They have $80 billion in cash and infinite credit lines.
Jensen is already working on the next big thing because he knows data centers is short term. And consider this, a full
Not surprised (Score:4, Insightful)
Given that they have optimized the crap out of the software they probably feel like they have done everything short of going assembly. But if you're going assembly you might as well just full send and make the hardware itself then you own the assembly.
It will be interesting to see where they go with this.
Re: (Score:3)
But you are correct that a new instruction set architecture is likely beneficial. It has nothing to do with programming languages though, it really has to do with the nature of the actions to be taken to solve a problem.
This is a gross oversimplification but I think it demonstrates the issue. Common CPUs are sort of designed to do what spreadsheets and word processors need. We designe
Re: (Score:2)
DeepSeek is designing a tensor processor. The instruction set is very simple. Mostly just lots and lots of parallelized low-precision multiply-and-accumulate.
Re: (Score:2)
Exactly. Building GPUs is difficult. Building AI chips is much simpler without having to deal with all the graphics parts.
Re: (Score:2)
That's an AI job right there, rewrite my old man code as assembly
Re: (Score:2)
That's an AI job right there, rewrite my old man code as assembly
That's what you compiler has been doing for decades. ;-)
Re: (Score:2)
mmm yes , good point 8)
Though purposefully written in assembly , from scratch, I always believed that the best chance of being the most efficient and can leverage hardware specifics? I've known game coders to write stuff in assembly, for example.
A lot the AI story is solving problems already solved.
Re: (Score:2)
Lithography (Score:5, Informative)
Buried at the end of the article hints at the real story:
Designing a competitive AI chip typically takes years and significant capital. Manufacturing poses another hurdle as the U.S. bans Chinese designers from accessing the most advanced overseas foundries, while separate U.S. curbs have cut China's access to high-bandwidth memory, a component critical to AI inference chips.
Not having access to ASML's lithography [wikipedia.org] puts any of these efforts at a pretty serious disadvantage. So companies making designs isn't really the story to watch but the outcome of Chinese lithography experiments [reuters.com].
Re: (Score:2, Informative)
For sure, Chinese chip making success or failure will be the real story.
A bunch of people are going to jump in now and say "derp see sanctions and tariffs ... have driven them toward domestic development"
No, this is not the case. The CCP wants a mono-polar world where they enjoy total hegemony. They have had their eye on the prize for two decades at least now. If you look at PRC history as far as 'friend factories' and collaborative efforts with soviet-bloc nations, and other technology transfer partnershi
Re: (Score:3)
Re: (Score:2)
I have not read the book but the argument sounds it could only come from a borderline retard.
What nation in all of history, under any political or economic system of organization has been immune to outside forces?
If you want to make the argument the US has been an aggressor, you're wrong communism was expansionist and we were reacting, you can but that isn't relevant.
I mean I could argue that 18th century monarchies failed because they were oerverthrown, invaded, corrupted, perverted, subverted, destabilize
Re: (Score:2)
AI is a gift for the Chinese. Even if their DRAM and AI training chips and GPUs aren't top tier, because of AI making everything else unobtainable, people will buy it.
The high prices mean they don't have to have small margins either, so there is more to invest back into R&D.
Re:Lithography (Score:4, Insightful)
Not having access to ASML's lithography puts any of these efforts at a pretty serious disadvantage. So companies making designs isn't really the story to watch but the outcome of Chinese lithography experiments
Or industrial espionage.
Or offering joint ventures to get the tech into the country and the workforce trained up.
Re: (Score:3)
Industrial espionage may help you with EUV lithography, but it sure isn't the main part of the problem.
(Actually, from what I've read recently, Chinese companies are going 3D before they get to EUV levels. That's another way around the problem. But you need to deal with a worse heat problem.)
Re: (Score:2)
They can probably live with lower efficiency because they have so much cheap energy.
Re: (Score:2)
Not sure how far industrial espionage gets you here.
There's nothing fundamentally secret in lithography. We know how it works. The secret sauce is the experience, processes, know-how, highly specialised suppliers etc. The practical complexities are immense. Given the security incidents ASML already had, I'd make a bet that a large amount of the valuable data is already in China. But the supply chain and the engineering don't live on paper and are not easily duplicated.
Re: (Score:2)
I guess you could frame it like that, but knowing how it works doesn't help if you don't have the know-how to build it.
How do you increase your EUV power from 100W to 1000W (this took ASML years to figure out)?
How do you make mirrors smooth to within a single atom deviation?
How do you make chemical etch pure to the parts-per-trillion level?
No doubt the Chinese will figure these things out by themselves if they have to, but I'm sure there are a few secrets to be had that would speed that up!
Re: (Score:2)
Probably.
But the mirrors, for example, are made by Zeiss. Lots of parts are from the supply chain and ASML doesn't even HAVE their secrets.
You are right that some industrial espionage could be useful. I just say it's not as useful as most people would assume (which is basically the Hollywood plot of "steal this secret and we can copy it"). Nah. You can find on the Internet how a nuclear bomb is made. But it takes a lot more than a print-out to actually make one, and a couple of those steps are genuinely har
Re: (Score:2)
That's irrelevant. Sure China has been blocked from buying ASML EUV machines, but they are doing just fine with DUV and companies like SMIC (Chinese semicondustor fab cf TSMC) have pushed it to ~7nm node size.
The Chinese are making their own AI accelerators, such as Huawei's Ascend series, which DeepSeek are using, but just like OpenAI making their own chips to avoid the NVIDIA tax, DeepSeek now want to make their own presumably at least in part to avoid the Huawei tax (as well as perhaps to gain even great
Re: (Score:2)
EUV would be nice to have, but at the end of the day it all comes down to cost. Without EUV your chips will be larger and slower, so you'll need more of them (more expensive) to build a cluster of the same power.
The cost of serving (not price to you) an AI model is mostly hardware depreciation cost, not operating cost, with the accelerator chip cost being a large part of that.
NVIDIA's H100 chip costs $25-40K
Huawei's comparable Ascend 950PR costs $7-16K
As you can see, lack of EUV is not stopping the Chinese
Re: (Score:2)
"Not having access to ASML's lithography [wikipedia.org] puts any of these efforts at a pretty serious disadvantage."
From the summary: "The chip is designed for inference"
Inference is more constrained by memory latency and capacity than by compute. You don't need EUV tools for DRAM (yet) and CXMT and YMTC have capacity. You can get pretty far on a mature logic node if you can pile on the memory.
Don't be too excited (Score:3)
Don't be too excited. It will only work in a limited set of servers and not with the standard AI frameworks. So even if you get the card, you probably can't run it. If you get the full server, you still need to rewrite your code. Unfortunately this isn't just a Nvidia clone you plug in and have fun with.
Re: (Score:2)
This isn't intended for consumer use at all - they are just doing the same as all the AI companies and trying to lower their inference cost by not paying a profit margin to someone else (in this case Huawei).