* **China’s AI Chip Revolution:** DeepSeek V3 is now supported on Huawei’s Ascend 910c, which can perform both training and inference—challenging Nvidia’s dominance with a competitive performance (60% of the H100 in standard benchmarks, with room for further optimization via hand-written CUNN kernels).
* **Ecosystem Disruption:** Huawei’s approach, featuring its own PyTorch repository that easily ports CUDA-like functionality to its CUNN framework, is set to undermine Nvidia’s traditional moat built on CUDA and its extensive software ecosystem.
* **Strategic Shift Away from U.S. Hardware:** With DeepSeek’s team actively optimizing its models for Huawei chips, Chinese innovators are reducing their dependency on export-restricted Nvidia GPUs, thereby significantly lowering operational costs and broadening access for previously sanctioned regions.
* **Implications for the Global Race to AGI:** These innovations not only disrupt existing market leaders but also signal a potential shift in global AI leadership, as a more diverse array of actors gains the computation
* **Once the new Huawei chip is available for testing, it might cause another huge stock market sell off.**
It seems Deepseek is now hosted on these new chips. Sure, maybe they are not more than 60% of H100, but give it 2-3 more years and they might catch up (just like Deepseek did).
Yogsothoz on
Necessity is the mother of innovation and boy did they innovate.
Shougee369 on
what about efficiency? does it matter for these kind of cards?
Responsible-Laugh590 on
Yes give it 2-3 more years while nvidia takes a nap 😴
danielv123 on
Important to note that the 910c is produced on smic 7nm, so they have managed to get their chip industry entirely domestic. It also has 53b transistors compared to 80b for the h100 so efficiency is up there.
Working directly in pytorch with 1 line to convert cuda to their language means Nvidia might not have a significant moat after all.
OmmmShantiOm on
Between this, Deepseek’s advances, and China investing 1 trillion yuans into AI, the race to AGI is really heating up. It seems we have hit a point of no return. No one is really discussing AI safety anymore, even though it could be a bigger existential threat than nuclear weapons. If the singularity is near, this is the most interesting time in human history. It may also spells the end of human history.
roarti on
It’s not only about hardware, it’s also about software, and CUDA is the standard currently. It would be great to have some competition because NVidias prices are ludicrous, but it’s hard to see for a lot of applications that rely on CUDA.
TLDR: China’s recent advancements in semiconductor technology, particularly with Huawei’s new 7nm chip, have surprised industry experts and raised questions about the effectiveness of US export restrictions. This breakthrough indicates that China is rapidly catching up in chip technology, potentially reshaping the global semiconductor landscape.
The two analysts in the podcast are super bullish on China, they say locking china out of current tech tree will only force China branching out to other tech pathways and they may develop their own tech indepedent of current chip industry. I think it is gradually happening right now.
Auran82 on
I think we’re going to look back and see AI to be like digital asbestos, an amazing, breakthrough product, with endless possibilities jammed into everything without considering the potential risks.
Giant_leaps on
The yield on these chips is probably not only very low but the chips themselves at 7nm are probably not as good or effective as nvdia chips. AMD have chips that are just as powerful but there isn’t as much demand because NVDA chips are so good that they become cheaper to operate
RyuzakiLawliet123 on
I’m not an expert by any means, but I like following industry developments, and from what Google tells me, the H100 was introduced in the third quarter of 2022, and there has been H200, B100 and B200 since. So would that mean that they’re 3 generations behind? Is this achievement all that impressive still? I’d appreciate it if someone could clarify, thanks!
11 Comments
Summary:
* **China’s AI Chip Revolution:** DeepSeek V3 is now supported on Huawei’s Ascend 910c, which can perform both training and inference—challenging Nvidia’s dominance with a competitive performance (60% of the H100 in standard benchmarks, with room for further optimization via hand-written CUNN kernels).
* **Ecosystem Disruption:** Huawei’s approach, featuring its own PyTorch repository that easily ports CUDA-like functionality to its CUNN framework, is set to undermine Nvidia’s traditional moat built on CUDA and its extensive software ecosystem.
* **Strategic Shift Away from U.S. Hardware:** With DeepSeek’s team actively optimizing its models for Huawei chips, Chinese innovators are reducing their dependency on export-restricted Nvidia GPUs, thereby significantly lowering operational costs and broadening access for previously sanctioned regions.
* **Implications for the Global Race to AGI:** These innovations not only disrupt existing market leaders but also signal a potential shift in global AI leadership, as a more diverse array of actors gains the computation
* **Once the new Huawei chip is available for testing, it might cause another huge stock market sell off.**
It seems Deepseek is now hosted on these new chips. Sure, maybe they are not more than 60% of H100, but give it 2-3 more years and they might catch up (just like Deepseek did).
Necessity is the mother of innovation and boy did they innovate.
what about efficiency? does it matter for these kind of cards?
Yes give it 2-3 more years while nvidia takes a nap 😴
Important to note that the 910c is produced on smic 7nm, so they have managed to get their chip industry entirely domestic. It also has 53b transistors compared to 80b for the h100 so efficiency is up there.
Working directly in pytorch with 1 line to convert cuda to their language means Nvidia might not have a significant moat after all.
Between this, Deepseek’s advances, and China investing 1 trillion yuans into AI, the race to AGI is really heating up. It seems we have hit a point of no return. No one is really discussing AI safety anymore, even though it could be a bigger existential threat than nuclear weapons. If the singularity is near, this is the most interesting time in human history. It may also spells the end of human history.
It’s not only about hardware, it’s also about software, and CUDA is the standard currently. It would be great to have some competition because NVidias prices are ludicrous, but it’s hard to see for a lot of applications that rely on CUDA.
Bloomberg had a podcast in 2023 when Huawei made a new mobile chip, the video is [here](https://youtu.be/pi2aTHCCV9U?si=nCaf1sX3VyFPpEkA). Here is an AI summary of the podcast:
TLDR: China’s recent advancements in semiconductor technology, particularly with Huawei’s new 7nm chip, have surprised industry experts and raised questions about the effectiveness of US export restrictions. This breakthrough indicates that China is rapidly catching up in chip technology, potentially reshaping the global semiconductor landscape.
The two analysts in the podcast are super bullish on China, they say locking china out of current tech tree will only force China branching out to other tech pathways and they may develop their own tech indepedent of current chip industry. I think it is gradually happening right now.
I think we’re going to look back and see AI to be like digital asbestos, an amazing, breakthrough product, with endless possibilities jammed into everything without considering the potential risks.
The yield on these chips is probably not only very low but the chips themselves at 7nm are probably not as good or effective as nvdia chips. AMD have chips that are just as powerful but there isn’t as much demand because NVDA chips are so good that they become cheaper to operate
I’m not an expert by any means, but I like following industry developments, and from what Google tells me, the H100 was introduced in the third quarter of 2022, and there has been H200, B100 and B200 since. So would that mean that they’re 3 generations behind? Is this achievement all that impressive still? I’d appreciate it if someone could clarify, thanks!