←back to thread

Alibaba Cloud says it cut Nvidia AI GPU use by 82% with new pooling system

(www.tomshardware.com)
521 points hd4 | 5 comments | 20 Oct 25 12:31 UTC | HN request time: 0s | source
Show context
hunglee2 ◴[20 Oct 25 12:59 UTC] No.45643396[source]▶
The US attempt to slow down China's technological development succeeds on the basis of preventing China from directly following the same path, but may backfire in the sense it forces innovation by China in a different direction. The overall outcome for us all may be increase efficiency as a result of this forced innovation, especially if Chinese companies continue to open source their advances, so we may in the end have reason to thank the US for their civilisational gate keeping
replies(17): >>45643584 #>>45643614 #>>45643618 #>>45643770 #>>45643876 #>>45644337 #>>45644641 #>>45644671 #>>45644907 #>>45645384 #>>45645721 #>>45646056 #>>45646138 #>>45648814 #>>45651479 #>>45651810 #>>45663019 #
lesuorac ◴[20 Oct 25 16:25 UTC] No.45645721[source]▶
The US isn't slowing China anymore.

China has an import ban on chips [1] so its irrelevant what the US does.

[1]: https://www.cnbc.com/2025/09/17/nvidia-ceo-disappointed-afte...

replies(3): >>45645872 #>>45645935 #>>45645986 #
xadhominemx ◴[20 Oct 25 16:42 UTC] No.45645935[source]▶
The US is certainly slowing down China considerably. China would certainly not have an import ban on Blackwell GPUs if they were made available. And upstream, the ban on EUV and other high end semiconductor production equipment has severely limited china’s capacity to produce logic and DRAM (including HBM).
replies(1): >>45652945 #
FooBarWidget ◴[21 Oct 25 06:09 UTC] No.45652945[source]▶
You severely underestimate what they can do with alternative tech paths. You don't have to chase nanometers for good AI system outcomes. Their current, and very viable, strategy is to build a ton of slower chips and to pump in a humongous amount of power. And to optimize the software stack, e.g. more efficient architectures. Unlike the west, they have a lot of cheap power (think solar panels in the desert) and excellent transmission. It also means they'll have to innovate on powder delivery and cooling systems to handle that sort of scale, but that's still easier than building EUV. Huawei has already done it with their phone from last year: they put in a more power-hungry chip, but they innovated hugely woth passive heat dissipation and bigger batteries, so the end product is still something consumers want. And with these Chinese AI models you're already seeing how they're reducing costs so they can run more on fewer chips.
replies(1): >>45658660 #
1. xadhominemx ◴[21 Oct 25 17:28 UTC] No.45658660[source]▶
There are workarounds and paths forward, but they are definitely being slowed down. I know a lot about this topic.
replies(1): >>45665688 #
2. FooBarWidget ◴[22 Oct 25 06:53 UTC] No.45665688[source]▶
That sounds like you are moving the goalpost: winning through a narrow definition of "slowing down". Defining "slowing down" as "not having access to western semiconductor ecosystem or 2nm chips" instead of "can they build economically viable AI systems/phones that achieve user objectives satisfactorily, and at scale" does not help your country/bloc with any meaningful outcomes, and merely serves to score Internet points to feel good.
replies(1): >>45675281 #
3. xadhominemx ◴[22 Oct 25 21:20 UTC] No.45675281[source]▶
No, we have certainly slowed down their progress in AI with export controls on GPUs and upstream. China would have better AI models but for those export controls.
replies(1): >>45680879 #
4. FooBarWidget ◴[23 Oct 25 12:04 UTC] No.45680879{3}[source]▶
"Better" in what way? Ultra-large frontier model performance? China doesn't need that to achieve strategic objectives, like increased productivity and automation, advanced weaponry, advanced manufacturing, user adoption and deployment at scale.

You're narrowing your claim into something you can defend but is strategically hollow. Advanced weaponry, major productivity improvements, R&D speed etc — at the end of day, those are the things the US bloc actually want to slow down. Ultra-large foundational modals was just (incorrectly) seen as the only way to achieve those objectives.

It's like you're arguing that Chinese fighters are inferior to western stealth fighters. That's true when you compare plane-by-plane on paper. And yet the Chinese airfighting system-as-a-whole was still able to down Rafaels with high precision and without being retaliated on, as shown by the India-Pakistan standoff a while ago. What's the point of arguing "we've slowed China's aircraft engine development speed" when they're still shooting down western jets?

replies(1): >>45689489 #
5. xadhominemx ◴[24 Oct 25 00:58 UTC] No.45689489{4}[source]▶
The Chinese definitely want their own frontier model. There is an enormous national effort behind building up the semiconductor manufacturing, data center infrastructure, and networking technologies required to compete with US frontier models. Because of the export controls, the semiconductor fab capacity required for the Chinese frontier models is at least 5x larger than the TSMC/Hynix capacity.