They did, but you’re not understanding what’s being said. The H20 is a crippled chip that NVIDIA made for the Chinese market because of US sanctions. This chip is about 50% slower than NVIDIA’s latest Blackwell B200 chip. China is saying the caught up to the crippled version. Which is true, they have not however caught up to the latest and greatest. They haven’t even caught up to the non-crippled version the H100 which the poster below mentions.
Do you have a source for english language benchmarks for the H20 versus say Huawei Ascend 910 (or honestly any latest Chinese compute GPU)? I am genuinely curious, independent benchmarks (at least in English) seem difficult to find. Searches in other languages that I do know lead to the same PR statements that are available in English.
I have done foreign language desk research (high level stuff, mostly looking for very specific things), but that was for a big paid project where we needed the data and I am not that interested in this topic. 😀
The impression I get (not a professional in this area, so maybe I don’t know what I am talking about) is that this a pure brute force approach for gaining performance at the rack system level.
It would be interesting to see direct benchmarks for H20 verses Ascend 910C across the different precision formats.
Ultimately what I would take from all of this is on hardware China is close. Not at NVIDIA chips yet, but without software optimization, it doesn’t matter how close the hardware is. Imo what China is doing right now is to force a criticality in number of users to find bugs and optimize the software. It’s literally the only way they can catch up. What’s hilarious is Lutnick literally told them this.
And to prove my point, just look at MooreThread S80. When it first released it had the hardware but ran so poorly it ran like a GTX1030. But with software optimization, it could run wukong at 48fps at 4k. Without the users, China can’t improve. So now they’re forcing people to be users.
They did, but you’re not understanding what’s being said. The H20 is a crippled chip that NVIDIA made for the Chinese market because of US sanctions. This chip is about 50% slower than NVIDIA’s latest Blackwell B200 chip. China is saying the caught up to the crippled version. Which is true, they have not however caught up to the latest and greatest. They haven’t even caught up to the non-crippled version the H100 which the poster below mentions.
Do you have a source for english language benchmarks for the H20 versus say Huawei Ascend 910 (or honestly any latest Chinese compute GPU)? I am genuinely curious, independent benchmarks (at least in English) seem difficult to find. Searches in other languages that I do know lead to the same PR statements that are available in English.
I have done foreign language desk research (high level stuff, mostly looking for very specific things), but that was for a big paid project where we needed the data and I am not that interested in this topic. 😀
Here’s an article about what I am saying. I don’t think there are benchmarks though.
https://www.theregister.com/2025/07/29/huawei_rackscale_boogeyman/#%3A~%3Atext=While+the+H20+still+holds+a+narrow+advantage%2Cat+least+so+far+as+inference+is+concerned.
The impression I get (not a professional in this area, so maybe I don’t know what I am talking about) is that this a pure brute force approach for gaining performance at the rack system level.
It would be interesting to see direct benchmarks for H20 verses Ascend 910C across the different precision formats.
It’s actually funny that you’re asking that. This is exactly Huawei’s solution.
https://semianalysis.com/2025/04/16/huawei-ai-cloudmatrix-384-chinas-answer-to-nvidia-gb200-nvl72/
Thank you! Will need to read it in more depth.
Ultimately what I would take from all of this is on hardware China is close. Not at NVIDIA chips yet, but without software optimization, it doesn’t matter how close the hardware is. Imo what China is doing right now is to force a criticality in number of users to find bugs and optimize the software. It’s literally the only way they can catch up. What’s hilarious is Lutnick literally told them this.
And to prove my point, just look at MooreThread S80. When it first released it had the hardware but ran so poorly it ran like a GTX1030. But with software optimization, it could run wukong at 48fps at 4k. Without the users, China can’t improve. So now they’re forcing people to be users.