Tencent Stockpiled Nvidia AI Chips for 'a Couple of Generations' (bloomberg.com) 23
Tencent dismissed concerns that US export controls will constrain its AI development capabilities, at least for the foreseeable future. From a report: The Shenzhen-based company has stockpiled Nvidia's H800 artificial intelligence accelerators, enough to develop its proprietary Hunyuan AI model for at least another couple of generations, President Martin Lau said on an analyst call after earnings on Wednesday. "Right now we actually have one of the largest inventories of AI chips in China among all the players," Lau said. "We were the first to put in orders for H800 and that allows us to have a pretty good inventory of H800 chips. So we have enough chips to continue our development."
The Biden administration in October escalated export controls on AI semiconductors heading to China, depriving the Asian nation from access to a broad range of the world's best AI-training hardware. The purpose of the measures is to prevent China's military from obtaining the advanced technology, Washington argues, but they're also making business harder for the country's private sector. Elsewhere in China, AI industry veteran Kai-Fu Lee's unicorn startup 01.AI has been buying up the Nvidia chips it needs to develop its own foundation models, with Lee saying the company has enough semiconductors for the next 18 months.
The Biden administration in October escalated export controls on AI semiconductors heading to China, depriving the Asian nation from access to a broad range of the world's best AI-training hardware. The purpose of the measures is to prevent China's military from obtaining the advanced technology, Washington argues, but they're also making business harder for the country's private sector. Elsewhere in China, AI industry veteran Kai-Fu Lee's unicorn startup 01.AI has been buying up the Nvidia chips it needs to develop its own foundation models, with Lee saying the company has enough semiconductors for the next 18 months.
Tencent's Doomsday Prepping Guide (Score:2)
It's not about how many chips you have; it's about how many you can actually use without setting your data center on fire.
Re: (Score:2)
Remember the good old days when private Chinese companies weren't just government fronts?
Re:Tencent's Doomsday Prepping Guide (Score:4, Informative)
No, I really don't.
Re: (Score:2)
As are anonymous cowards.
Re: (Score:1)
They mostly come at ... mostly
Re: (Score:2)
Not sure how many octogenarians there are reading slashdot. I think there is at least a few.
Stockpiling CPUs isn't a good idea... (Score:2)
But given the acceleration that both sofware and hardware for AI are achieving [ieee.org], Tencent's stockpile will be obsolete in just a few months...
Re: (Score:2)
Tencent to seek domestic AI training chip supplies after latest US curbs
Chinese tech group Tencent Holdings (0700.HK) said on Wednesday it will look for domestic sources for AI training chips following new moves by the United States to restrict chip supplies to China.
Tencent president Martin Lau said on a call with analysts that the White House's decision last month to ban more AI chips from export to China will force the company to use its exi
Re: (Score:2)
they only need the stock until they can produce their own new chips. they will laugh at these sanctions just like they did at the ones for cpus with their brand new kirin which was also "impossible".
Re: (Score:3)
they only need the stock until they can produce their own new chips. they will laugh at these sanctions just like they did at the ones for cpus with their brand new kirin which was also "impossible".
With CPUs to can adhere to a known architecture and the compilers will do decent, even if they can't use the chip specific optimizations.
The issue with GPUs is that the driver is a lot more important as well as the integration with the toolkit.
So even if you build a really nice GPU and write a really good driver you'll need to write your own ML libraries to take advantage of it.
Tencent is probably big enough to do that for their own in-house models. But they're not going to be be able to build support for t
Re: (Score:2)
Keep in mind that when you say "the libraries that everyone else uses", you really mean "the libraries that Westerners use". Things are different in China. Chinese devs are more likely to be using Tencent libraries, with documentation in Chinese.
Same thing has happened with electronics. There are all these parts that only have Chinese datasheets and only get used by Chinese engineers, and millions, even billions of them are being made.
Re: (Score:2)
Keep in mind that when you say "the libraries that everyone else uses", you really mean "the libraries that Westerners use". Things are different in China. Chinese devs are more likely to be using Tencent libraries, with documentation in Chinese.
Same thing has happened with electronics. There are all these parts that only have Chinese datasheets and only get used by Chinese engineers, and millions, even billions of them are being made.
The difference with electronics is that you need to make things clean and precise, but the complexity isn't that high. That's why it got outsourced so thoroughly.
Assuming you're bilingual in Chinese and English, which are you going to use to build your ML, Tensorflow, PyTorch, or whatever Tencent is building?
Those libraries might work fine for whatever Tencent needs. But if they're building LLMs and you need an LSTMs for regression models you their stuff is going to work as well as the libraries that Wester
Did this artificially bloat NVidia sales numbers? (Score:2)
Attempting to prevent... (Score:2)
...China from acquiring tech is futile and will backfire
They will get what they need and may someday develop stuff we need, and refuse to sell it to us in retaliation
Re: (Score:2)
China is way behind in semiconductor manufacturing which is why they stockpiled the chips. Most of what the US gets from China is raw materials which can be obtained elsewhere or sourced domestically. We are under no obligation to provide them with our latest tech.
Re: (Score:1)
Re: (Score:2)
It takes a long time to reverse engineer advanced technology, and by that time it is no longer advanced. China is at least 10 years behind on semiconductors and unlikely to catch up.
Re: (Score:2)
And there's no guarantee that methodologies will even be the same by them. There's for example Tesla's insanely-parallel Dojo architecture, which is finally starting to scale. Or we might go to neuromorphic hardware, where you're outright working in analog.
If all you can do is copy, well, good luck.
The horse didn't bolt (Score:2)
It's purposefully standing just outside the barn mocking you while you close the door.