As the China-US chip escalates, Chinese companies have started finding alternatives to NVIDIA’s AI GPUs which were previously their main component for data center and cloud ecosystems.
NVIDIA Out From China – Several AI Alternatives To Take Benefit From The Huge Domestic Market, One GPU Maker Calls AI Op “A Sea of Stars”
Last month, The US government hit China with a new series of bans on its AI efforts which blocked NVIDIA from exporting its latest Hopper and Ampere chips. This new ban was an interesting one as it stopped NVIDIA from selling its China-exclusive A800 & H800 AI GPUs too which were made just to comply with the previously placed bans and had limited functionality compared to the full chips that were sold in the US and other regions. The ban also affected other countries that either have strong ties with China or are involved in selling their AI GPU supplies to Chinese customers through unofficial and black market channels.
To sum things up:
- The export control regulations restrict NVIDIA, AMD, & Intel from selling AI chips to China with compute capabilities of over 4000 TOPS (AI Compute).
- More than 40 countries have been blocked from exporting AI GPUs from the aforementioned companies to prevent China from acquiring chips from them.
- 21 countries and regions have imposed restrictions on selling advanced manufacturing equipment to China to prevent them from accessing chip manufacturing technology below 14nm.
- Chinese startups (Moore Threads, BirenTech) have also been added to the import list to prevent domestic companies from producing high-end “Domestic” AI solutions.
This ban has prompted NVIDIA to cancel AI GPU orders worth over $5 Billion US from various Chinese companies but the GPU giant will have no trouble in new customers due to massive demands for these chips and no real alternative at the moment that competes with Hopper or Ampere. The company has also started to reallocate its A800 & H800 AI GPUs to the global AI market as seen just a few days ago.
So with NVIDIA out of the equation, what real AI alternatives does the Chinese market have to serve its domestic market? Well, the answer was provided in a recent seminar in China where various AI & tech companies announced their developments for the domestic market and the solutions on offer that will serve the Chinese AI segment. The companies come in three tiers which include:
1st Tier “Technology” Companies:
2nd Tier “Enterprise” Companies:
3rd Tier “Entrepreneurial” Companies:
- Moore Threads
- Tianshu Zhixin
All of these companies are either offering or have plans to produce AI chips that will serve as NVIDIA alternatives in the Chinese domestic market. Their products are positioned in three main tiers which include high-end, mid-high-end, and mid-lower-end. The flagship chip is the Huawei Ascent 910 which is made on a 7nm process node from TSMC and directly competes with NVIDIA A100 and A800 GPUs. The chip can reach around 80% Inference and 70% training performance of the NVIDIA A100 accelerator but those are theoretical metrics and the actual performance is still a bit lacking compared to NVIIA’s solutions.
There’s also the Alibaba Hanguang 800 which is based on a TSMC 12nm node and is positioned against NVIDIA’s P4 while the Tencent Zixiao is positioned against the NVIDIA A10. The last two chips target the mid-lower-end market. Other chips include the Intel Gaudi, Jing Jiawei JM9, Biren BR100, Moore Threads MTT S3000, and a range of other AI accelerators from Chinese companies but one thing that’s common amongst all is that they are all based on TSMC’s process technologies ranging from 12nm and 7nm.
Intel’s Gaudi2 accelerators were turning up to be quite the alternative to NVIDIA’s AI GPUs in China prior to the ban. Meanwhile, Birentech, a promising AI startup that was going to offer performance rivaling NVIDIA’s top AI GPUs, was not allowed to utilize TSMC’s manufacturing capabilities to make its next-gen chips.
Meanwhile, the CEO of Moore Threads has stated that while China has been hit hard with the recent US bans for AI chips, this is not to be taken as a dark moment but rather to be seen as a “Sea of Stars” as it offers domestic chipmakers an opportunity to showcase their prowess in AI chip development and offer solutions in their own market. The company is also on the path to offering its next-gen GPUs in the coming years.
Because of the 1017 incident, the entire domestic GPU/AI chip industry has been hit hard. At this point in time when challenges and opportunities coexist, what I want to say is that there is no “darkest moment” for Chinese GPUs, only a sea of stars.
Moore Threads has only one mission from beginning to end: to build China’s best full-featured GPU. We will carry this undertaking to the end, and nothing will affect our determination to continue. The more difficult it is, the faster we must move forward.
Before starting the next three-year plan, I would like to emphasize a few points:
- Accelerate independent research and development and innovation. At present, our number of authorized patents has temporarily led the country, but we still need to work harder to reach our goal
- Polish high-quality products. The software team must further optimize product performance at 2x speed, enhance product competitiveness, and create higher value for users
- Promote application implementation. Ecology is the moat of GPU. DevRel and GTM teams must continue to expand their circle of friends and let MUSA support more industry applications
- Build an efficient team. Strengthen organizational management, focus on the research and development of full-featured GPU core technology, and maintain team efficiency and agility
Jams Zhang (CEO of Moore Threads) via MyDrivers
It will be interesting to see how the market evolves since these companies are now the only AI alternative for China’s huge domestic market which hungers for more chips. It is likely that the restrictions on China will continue to get intense under the current Biden administration & which will only prompt Chinese AI companies to pool in all their efforts to create fast and competitive chips that rival the likes of NVIDIA, AMD, and Intel.