Argomenti trattati
Introduction to Huawei’s Ascend 910D
Huawei is stepping up its game in the AI processor arena with the upcoming Ascend 910D. This new chip is designed to outperform Nvidia’s H100, a formidable competitor in the AI landscape. While the Ascend 910D may initially fall short in raw chip-to-chip comparisons against Nvidia’s Blackwell series, its innovative architecture—featuring clusters of processors—could provide the competitive edge needed to challenge Nvidia’s stronghold.
Performance expectations and testing
The road to exceeding Nvidia’s performance benchmarks is fraught with challenges. Huawei has already begun engaging local companies to evaluate the Ascend 910D’s capabilities. Initial testing of this advanced processor is anticipated to commence by late May. This strategic move aims to ensure that the chip meets both performance standards and deployment requirements set by the industry.
In parallel, Huawei is preparing for large-scale shipments of its dual-chiplet Ascend 910C processors, expected to reach Chinese customers shortly. These chips were primarily manufactured by TSMC for a third party, indicating Huawei’s reliance on external partners for cutting-edge technology. The production method for the Ascend 910D remains uncertain—whether it will rely on China’s SMIC or find a workaround to U.S. sanctions is yet to be determined.
Challenges in achieving competitive performance
Matching Nvidia’s H100 performance will be a significant hurdle for Huawei. The Ascend 910C currently delivers approximately 780 BF16 TFLOPS, while the H100 achieves around 2,000 BF16 TFLOPS. To bridge this gap, Huawei must consider redesigning the internal architecture of the Ascend 910D, potentially increasing the number of compute chiplets to enhance its processing power.
The competitive landscape is further complicated by Nvidia’s upcoming Rubin GPUs, set for release in 2026. These GPUs are expected to leverage advanced fabrication processes that could leave Huawei at a disadvantage without access to similar technology.
Huawei’s strategy with CloudMatrix systems
This year, Huawei introduced the CloudMatrix 384 system, which integrates 384 Ascend 910C processors. Early reports suggest that this system can outperform Nvidia’s GB200 NVL72 in specific workloads, although it does so at the expense of higher power consumption. This highlights a crucial aspect of AI processing: balancing performance with energy efficiency.
The effectiveness of these systems hinges on whether the interconnect technology can scale efficiently to accommodate the required number of processors, a critical factor in the overall success of Huawei’s strategy.
The future of AI processing in China
As the AI sector continues to evolve, Huawei’s Ascend 910D is poised to become a cornerstone for AI training within China. Given the strategic importance of AI technology, the power consumption of these processors may not be as significant a concern as their performance output. The number of units deployed could compensate for efficiency disparities, solidifying Huawei’s role in the domestic AI landscape.
However, Huawei’s success will largely depend on its ability to scale production capabilities, whether through domestic manufacturing or international partnerships. The geopolitical landscape and ongoing trade tensions will likely play a pivotal role in shaping Huawei’s future in AI processor development.
Conclusion
As Huawei navigates these challenges, the Ascend 910D represents a bold step forward in the quest for AI supremacy. The outcome of its efforts will not only determine Huawei’s position in the technology sector but also influence the broader global AI landscape.