Huawei’s HiSilicon Ascend 910C AI processor delivers 60% of the inference performance of Nvidia’s H100, according to recent findings by DeepSeek. This aging yet evolving chip could potentially reduce China’s dependence on Nvidia GPUs. The Ascend 910C demonstrates notable inference capability despite not being a training performance leader.
Researchers from DeepSeek have shown that the Ascend 910C can exceed expectations in inference performance, with further optimizations of CUNN kernels potentially enhancing its efficiency. DeepSeek’s support for Ascend processors and its PyTorch repository facilitate a smooth transition from CUDA to CUNN, making Huawei’s hardware more accessible for AI workflows. Huawei’s development faced challenges due to U.S. sanctions and limited access to TSMC’s leading-edge process technologies.
Despite this, Huawei, in collaboration with SMIC, has managed to develop a competitive chip akin to Nvidia’s A100 and H100 processors.
DeepSeek supports Huawei’s AI innovation
The Ascend 910C employs chiplet packaging and boasts a compute SoC with approximately 53 billion transistors, using SMIC’s 2nd Generation 7nm-class process technology known as N+2.
Although the Ascend 910C excels in inference tasks, Nvidia maintains a significant lead in AI training due to its well-established hardware and software ecosystem. DeepSeek’s Yuchen Jin highlighted that long-term training reliability remains a critical issue for Chinese processors, stemming from the two-decade development head start that Nvidia holds. Experts predict that as AI models continue to converge towards Transformer architectures, the significance of Nvidia’s software ecosystem may wane.
With continued optimization efforts from firms like DeepSeek, the dependency on Nvidia could diminish, presenting a cost-effective alternative for AI companies, particularly for inference tasks. However, to compete on a global scale, China must enhance the stability of its AI training infrastructure. Moving forward, the AI and semiconductor industries will closely watch how Huawei’s strategies evolve to address these challenges and opportunities.
Cameron is a highly regarded contributor in the rapidly evolving fields of artificial intelligence (AI) and machine learning. His articles delve into the theoretical underpinnings of AI, the practical applications of machine learning across industries, ethical considerations of autonomous systems, and the societal impacts of these disruptive technologies.
