China firms' AI breakthrough can meld GPUs from different brands into one training cluster — Baidu says new tech fuses thousands of GPUs together to help sidestep shortages

Baidu
(Image credit: Shutterstock)

Tech companies are increasing their pace in developing AI technologies. Baidu, China’s largest search engine, is often compared to Google and is pushing hard to gain a global lead. According to its CEO, Robin Li, the company has developed a system to combine GPUs from different vendors and use them as a single compute cluster for AI training. This was highlighted in its Q1 2024 Earnings Call transcript.

Li mentioned in the call that Baidu is transforming from an Internet-centric business to an AI-first business with its generative AI, ERNIE, becoming the new core of its products in 2025. However, we expect Baidu to encounter headwinds, especially as it’s a Chinese company, and the US severely restricts the export of advanced technologies to China, such as the latest-generation Nvidia, AMD, and Intel chips, which are crucial to AI development.

Latest Videos FromTom's Hardware
TOPICS
Jowi Morales
Contributing Writer

Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.

  • As Chinese President Xi Jinping said after the Netherlands blocked the export of ASML's lithography equipment,

    Is that what Chinese President said via that link ? This is funny ! :LOL:

    How to speed up a slow Mac ?
    https://www.macworld.com/article/668632/how-to-speed-up-a-mac.html
    The White House's attempts to stop technology transfers from the US and its allies may hinder China's scientific advancements in the short term.

    Sorry! Page not found.
    The page you're looking for has either been moved or removed from the site.
    Please try searching our site or start again on our homepage.



    Maybe try proof-reading articles BEFORE publishing ?
    Reply
  • bit_user
    The article said:
    The company can combat AI chip scarcity by combining GPUs from different vendors.
    ...
    If Baidu's claims are true, this is a massive development.
    How much is that going to help, really? I'm just not seeing a huge upside, here. If you pair a handful of Chinese GPUs with a Nvidia GPU that's 10x as fast, then the total benefit on training time won't add up to much. Also, for anyone building AI systems, they're dealing in aggregates and I'll bet they build systems with either all Nvidia or all AMD GPUs. It's probably much more the exception that they're down to just a couple boards of either kind, and if you were, you just build another all-AMD system (for instance).

    Furthermore, any time you don't have a high-speed fabric and have to rely on PCIe for interconnectivity, you're going to be at a significant disadvantage. The software overhead of abstracting each GPU API is going to add a little more, but I see it as neither a huge win nor a major impediment.

    The article said:
    If Li's claim is true, Baidu has achieved a brilliant technical breakthrough. This technology will allow the company to mix and match different GPUs
    On a purely technical level, I think it's less impressive than prior techniques for mixing & matching different GPUs in the same machine.
    https://www.anandtech.com/show/2844
    https://www.anandtech.com/show/4522/ecs-p67h2a-review-a-visit-back-to-lucids-hydra/8
    While searching for that, I found a project enabling disparate multi-GPU configurations for CFD:
    https://www.reddit.com/r/CFD/comments/107mzx5/the_fluidx3d_v20_multigpu_uptate_is_now_out_on/
    Sorta shows it's not quite the genius breakthrough the article claims.
    Reply
  • JTWrenn
    This has always been doable, the question is how large is the efficiency dropped by doing it.
    Reply
  • ThomasKinsley
    bit_user said:
    How much is that going to help, really? I'm just not seeing a huge upside, here. If you pair a handful of Chinese GPUs with a Nvidia GPU that's 10x as fast, then the total benefit on training time won't add up to much. Also, for anyone building AI systems, they're dealing in aggregates and I'll bet they build systems with either all Nvidia or all AMD GPUs. It's probably much more the exception that they're down to just a couple boards of either kind, and if you were, you just build another all-AMD system (for instance).

    Furthermore, any time you don't have a high-speed fabric and have to rely on PCIe for interconnectivity, you're going to be at a significant disadvantage. The software overhead of abstracting each GPU API is going to add a little more, but I see it as neither a huge win nor a major impediment.


    On a purely technical level, I think it's less impressive than prior techniques for mixing & matching different GPUs in the same machine.
    https://www.anandtech.com/show/2844https://www.anandtech.com/show/4522/ecs-p67h2a-review-a-visit-back-to-lucids-hydra/8
    While searching for that, I found a project enabling disparate multi-GPU configurations for CFD:
    https://www.reddit.com/r/CFD/comments/107mzx5/the_fluidx3d_v20_multigpu_uptate_is_now_out_on/
    Sorta shows it's not quite the genius breakthrough the article claims.
    I think you've nailed it. To me this sounds like SLI on steroids, but SLI suffered from overhead with just two cards that made it a losing proposition. I can't imagine how what the overhead would be with hundreds or thousands of GPUs in a cluster. While the current tech is not impressive on a technical level, I still think it would probably be useful for LLM training during a supply shortage from the sanctions.
    Reply
  • Pierce2623
    I’ll happily bet a hundred dollars that it’s literally just the recompilers made to run CUDA code on AMD or Intel just combined into one package.
    Reply
  • bit_user
    ThomasKinsley said:
    SLI suffered from overhead with just two cards that made it a losing proposition. I can't imagine how what the overhead would be with hundreds or thousands of GPUs in a cluster. While the current tech is not impressive on a technical level, I still think it would probably be useful for LLM training during a supply shortage from the sanctions.
    The communication patterns are very different for training vs. rendering. One of the first things the US sanctions targeted was communication bandwidth, because they knew how much that would impede training large-scale models. There are good reasons for Nvidia's focus on NVLink. AMD and Intel each have their own version, as well.
    Reply
  • Gillerer
    ThomasKinsley said:
    I think you've nailed it. To me this sounds like SLI on steroids, but SLI suffered from overhead with just two cards that made it a losing proposition. I can't imagine how what the overhead would be with hundreds or thousands of GPUs in a cluster. While the current tech is not impressive on a technical level, I still think it would probably be useful for LLM training during a supply shortage from the sanctions.

    SLI overhead affected gaming because the need for precise synchronization and low latency. (And what finally killed it was the impending move to low-level APIs such as DX12 and Vulkan, where it would have fallen on game developers to implement multi-GPU on each game.)

    Depending on the kind of compute workload you have, these might not be a factor at all. Multi-GPU still has applications for compute and rendering tasks (just not real-time rendering).
    Reply