Huawei is gearing up to introduce its latest AI accelerator, the Ascend 910C, which is expected to hit the market soon. The chip is already in production, and we now have insight into its anticipated capabilities.
As China’s AI industry undergoes a significant “technological” transformation, there’s a soaring demand for greater computing power, especially with models like DeepSeek’s R1 entering the scene. Among the Chinese AI chip producers, Huawei is definitely making waves. Their forthcoming Ascend 910C is set to intensify competition, notably posing a challenge to NVIDIA’s H100. Recent analysis by industry expert Lennart Heim has divulged intriguing details about the 910C, hinting that it could be a strong contender in the Chinese market.
Based on the information from the analysis and our own findings, the Ascend 910C seems to avoid the complex design seen with some NVIDIA chips. Instead, Huawei is opting for a more conventional approach by using two separate silicon interposers linked through an organic substrate, essentially doubling up on its Ascend 910B chips for enhanced performance. The expected result is impressive, with the 910C projected to hit 800 TFLOP/s at FP16 and offer up to 3.2 TB/s memory bandwidth, nearly matching NVIDIA’s H100 capabilities.
From a production standpoint, Huawei plans to leverage the 7nm process from both TSMC and SMIC. Before export restrictions took effect, Huawei secured significant 7nm orders with TSMC, ensuring a substantial portion of the chip’s production. Additionally, SMIC, China’s leading semiconductor company, has successfully developed its own 7nm process, likely contributing to Huawei’s ambitious goal of shipping “millions” of AI chips.
Considering China’s advancements in domestic AI technology, the Ascend 910C marks a significant milestone. While Huawei still lags behind global leaders in terms of pure computing power—often by a factor of 10 to 20—there are innovative solutions emerging within the country. Companies like DeepSeek have demonstrated alternative strategies for managing limited computing resources, hinting that the future of the AI race could be fiercer and more unpredictable than ever.
