Huawei is preparing to ship its Ascend 910C AI chip in volume next month, offering a crucial alternative for Chinese companies facing US restrictions on advanced semiconductors. The move comes as limitations on Nvidia’s high-end AI chips, like the H20, intensify the need for domestic solutions capable of handling large-scale AI training and inference tasks.
The Ascend 910C, essentially a doubled-up version of the earlier 910B, achieves performance comparable to Nvidia’s H100 by combining two processors. Huawei’s strategy focuses on scaling via high-speed interconnects and massive parallelization, rather than relying on cutting-edge process node technology. This is illustrated by their CloudMatrix 384 system, a rack-scale platform featuring 384 Ascend 910C chips.
The CloudMatrix architecture, as reported by SemiAnalysis, leverages 6,912 800G LPO optical transceivers to create an optical all-to-all mesh network. This reportedly allows the system to deliver around 300 petaFLOPs of BF16 compute power, potentially exceeding the performance of Nvidia’s GB200 NVL72. The CloudMatrix also demonstrates strengths in memory bandwidth and capacity.
While the ‘brute-force’ approach of the Ascend 910C may result in lower power efficiency compared to Nvidia’s GB200, it provides critical AI infrastructure. Reports suggest the chips are manufactured by SMIC using its 7nm N+2 process, though yield levels remain a concern. Controversy surrounds some units reportedly produced by TSMC for Chinese firm Sophgo, a claim Huawei denies. The US Commerce Department is investigating any links between TSMC and Sophgo.
According to Albright Stonebridge Group, the Ascend 910C is expected to become a preferred option for Chinese companies prioritizing large AI model development and inference capacity growth, especially given the US chip restrictions. Despite performance trade-offs, the Ascend 910C represents a broader trend of Chinese firms developing indigenous technologies to bolster supply chain security and mitigate geopolitical risks.