Huawei's next AI accelerator—the Ascend 910C—is entering production. It's China's best AI chip. Thanks to backdoor sourcing, we could easily see 1M H100-equiv this year. Here’s what we know about its performance and strategic implications. Spoiler: selectively competitive. 1/

Mar 11, 2025 · 3:51 PM UTC

44
396
37
2,271
The 910C is basically two co-packaged Ascend 910Bs, China's best current-gen accelerator. But there's a twist: most (potentially all) of these chips weren't produced domestically—they were illicitly procured from TSMC despite export controls. 2/
9
16
8
154
I'd expect the 910C to achieve ~800 TFLOP/s at FP16 and ~3.2 TB/s memory bandwidth. This makes it only ≈80% as performant as NVIDIA's previous-generation H100 (from 2022) while using 60% more logic die area. 3/
2
7
1
121
Unlike NVIDIA's advanced packaging in the B100/200 series, the 910C likely uses a less technically sophisticated approach with two separate silicon interposers connected by an organic substrate. 4/
1
6
1
89
This could result in 10-20x less die-to-die bandwidth compared to NVIDIA's solution. This needs to be overcome by engineering. If the bandwidth is that low, it's not really one chip, and engineers using that chip need to take it into account. 5/
1
5
2
85
The technical gap is substantial: compared to NVIDIA's B200 that will go into data centers this year. The 910C has ~3x less computational performance, ~2.5x less memory bandwidth (assuming HBM2E which they've stockpiled; HBM3 also possible), and a lot more power-inefficient. 6/
1
7
1
90
Huawei likely illicitly got close to 3M Ascend dies (7nm) from TSMC (now fixed via foundry due diligence rule). They also stockpiled HBM2E memory from Samsung (also controlled but they stockpiled before): enough for potentially 1.4M 910C accelerators. xcancel.com/Huang_Sihao/status/187… 7/
Sihao Huang @Huang_Sihao
Jan 15
In October 2024, it was widely reported that TSMC had fabricated a large number of export-controlled AI accelerator dies for Huawei. The new foundries rules are aimed at preventing Chinese firms from circumventing export controls in a similar manner. 1/
1
8
2
80
In addition, @Gregory_C_Allen just shared some speculations on their own advanced production capacity. They should be able to produce 910B and 910C dies at the 7nm node. xcancel.com/Gregory_C_Allen/status… 8/
Gregory C. Allen
 @Gregory_C_Allen
Mar 7
I am extremely proud to share that my new report "DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race" has just been published by CSIS! This report is chock full of previously not-public information and analysis connecting dozens of data points to assess where DeepSeek's achievements do and do not change the state of U.S.-China AI competition. Tldr: it's not a psyop, and it is a big deal. BUT Huawei's recent success in two areas matters most. 1) getting TSMC to manufacture boatloads of Ascend AI chips, and 2) Smuggling advanced chip-making equipment to SMIC Over the long term, the biggest risk of DeepSeek is that the community of Chinese open source enthusiasts get interested in running DeepSeek models on Huawei CANN software, which could pose a threat to Nvidia's CUDA moat. Right now DeepSeek isn't interested in Ascends because they're lousy chips with dreadful supporting software, but at least the software part could change in a few years. Please read the report and help spread the word! Link in next tweet
1
8
1
71
But we've yet to see a teardown of a 910B or 910C actually produced domestically (I think it's possible but expect the majority to come illegally from TSMC). 9/
1
3
1
56
While impressive, this still falls short of what the West produces, with at least 5x the number of chips in 2025 and 10-20x the computing power. The US compute advantage in total remains strong. 10/
2
4
1
64
Having 10x more compute is cool and a key strategic advantage, as I've argued before. But it's different if it's disbursed across many companies. China can centralize more easily than we do... that's a key thing to watch out for. 11/
1
3
2
65
This means China will be competitive in many domains. Expect competitive models and more gains especially from reasoning. However, the next pre-training generation might require new and bigger clusters needing tens of thousands of chips. 12/
1
3
1
55
Furthermore, to gain from those models, countries will want to deploy them to millions of users, or run a large number of AI agents autonomously, where total compute quantity still matters. That’s where we will see the impact of these controls. 13/
1
3
1
43
To summarize: Per-chip performance isn't impressive—achieving only 80% of the H100 with a 4 year delay. BUT, they can overcome it by clustering more chips given the substantial amount of illicit dies procured from TSMC (and potentially smaller amounts from SMIC). 14/
1
7
2
52
There will be competitive models from China—the talent and compute are there. This doesn't mean export controls failed; it's just critical to understand what China can deliver, what export controls allow, and what they do not. 15/
1
2
1
45
I've shared before all the complementary approaches we need—AI resilience, AI for defense, and more. Will write this all up together at some point to pre-empt another DeepSeek-style freakout. Thanks to @Huang_Sihao and others! 16/16
7
3
60