Huawei and DeepSeek Narrow AI Chip Ecosystem Gap with Nvidia

iconFLASH AI
Share
AI summary iconSummary
Huawei and DeepSeek are closing the gap with Nvidia's chip ecosystem, according to a recent daily market report by FLASH AI and SemiAnalysis. The report highlights DeepSeek's use of Huawei's Ascend 950PR chips and software updates as signs of progress. A fear and greed index for the AI sector may soon reflect this shift. DeepSeek plans to rely on Huawei's stack by 2027. Despite this, the Ascend 950PR still trails Nvidia's H100 in performance.

SemiAnalysis said in research published on Thursday that closer collaboration between Huawei Technologies and DeepSeek is narrowing the gap with Nvidia's chip hardware and software ecosystem. SemiAnalysis cited DeepSeek's recent software releases and increased experimentation with Huawei's Ascend 950PR chips as evidence that China's technological gap with the United States is shrinking. SemiAnalysis expects DeepSeek to depend mainly on Huawei's technology stack by 2027 or 2028. SemiAnalysis described Huawei's stack as mature and well developed, citing external and internal pressure to create an entirely domestic technology ecosystem in China.

DeepSeek released tools for the Ascend ecosystem weeks before the research appeared. The releases included six software modules corresponding to modules the Hangzhou start-up had previously built for Nvidia chips. DeepSeek also released an Ascend-compatible version of TileLang, a custom programming language that simplifies the development of high-performance kernels, essential code that operates chip hardware. DeepSeek said late last month that the Ascend version could support a more independent, domestically controlled software ecosystem for graphics processing units.

The research describes DeepSeek testing model-specific workloads within Ascend's hardware and software ecosystem. DeepSeek has also given Huawei early access to DeepSeek models. Huawei announced inference support from launch for major releases, including DeepSeek V4 and V4.1. Inference uses chips to run an AI model rather than train the model.

However, SemiAnalysis said Huawei's Ascend 950PR, unveiled in March, remains behind Nvidia's H100 in absolute performance. Nvidia started shipping the H100 in 2022. Huawei engineers' best Ascend kernels have reached 99.7% of the kernels' theoretical efficiency ceiling.

SemiAnalysis identified potential compatibility problems for hardware configurations beyond DeepSeek's particular supernode setups. Most optimization work so far has targeted large-scale supernodes. Problems remain for developers using Ascend hardware with standard x86 host processors rather than Huawei's native architectures. SemiAnalysis said other operators running DeepSeek models could reuse the kernels and deployment methods developed through the collaboration. SemiAnalysis said other laboratories could develop models using shared compiler capabilities. SemiAnalysis said improvements could accumulate if those users contribute fixes and optimizations.

Beijing has prioritized technological self-reliance amid US restrictions on China's access to advanced chips, including Nvidia products, and chipmaking equipment. Nvidia has become increasingly excluded from China's market. Nvidia's Jensen Huang has become more vocal about the restrictions' substantial effects on Nvidia and US technological competitiveness.

Huang spoke in a podcast aired in April, shortly before DeepSeek launched the V4 foundation model. Huang said DeepSeek optimizing models for training and inference on Huawei chips would be a terrible outcome for the United States. Huang claimed China would surpass the United States as a result. SemiAnalysis said software results and 950PR testing showed that DeepSeek's contribution, alongside Huawei's work, was reducing the ecosystem gap and beginning to realize Huang's feared scenario.

Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information. Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.