Chinese CPU Achieves End-to-End Support for DeepSeek AI Models

Beijing, China – In a significant stride for China’s domestic semiconductor industry, Phytium Information Technology Co., Ltd. announced yesterday that its Tengyun S5000C and Tengrui D3000 CPUs have successfully achieved end-to-end support for the DeepSeek series of large language models (LLMs). This breakthrough covers both data center and terminal scenarios, with locally deployed inference efficiency reportedly comparable to solutions based on foreign architectures.

The announcement, reported by IT Home, marks a crucial step towards reducing reliance on foreign technology in the rapidly evolving field of artificial intelligence.

DeepSeek-V2.5 236B Runs Stably on Phytium Platform

According to Phytium, the DeepSeek-V2.5 236B large model was successfully and stably operated on the Phytium Tengyun S5000C platform in 2024. The company claims that a configuration of the Tengyun S5000C paired with eight domestically produced AI cards achieves approximately 80% of the performance of comparable foreign platforms.

Following the release of DeepSeek-V3 and R1 versions, the Phytium technical team conducted local deployment and verification on both the Tengyun S5000C and Tengrui D3000 platforms, resulting in a comprehensive solution.

Data Center and Terminal Solutions

The solution offers flexibility for different deployment environments:

Data Center: Phytium’s Tengyun S5000C servers can be combined with various AI acceleration cards to deploy both distilled and full versions of DeepSeek. The Phytium AI Laboratory has tested the distilled version DeepSeek-R1-Distill-Llama-70B on a single-machine 8-card server platform, achieving an inference speed of 22 tokens per second. The full DeepSeek R1 671B model utilizes a distributed inference framework and can be successfully deployed on a four-machine, 32-card AI server cluster.
User Terminal: The Phytium Tengrui D3000-based desktop can locally deploy the DeepSeek-R1 7B distilled model, coupled with a domestic inference card, enabling AI PC capabilities.

Phytium CPUs: Key Specifications

The announcement highlights the specifications of the two CPUs at the heart of the solution:

Tengyun S5000C: Available in 64-core, 32-core, and 16-core configurations, the Tengyun S5000C is compatible with the ARMv8 instruction set and supports hardware virtualization. It is designed for a wide range of applications, including computing servers, storage servers, AI servers, high-end network security, industry-grade business host servers, and large internet data centers.
Tengrui D3000: The Tengrui D3000 integrates eight of Phytium’s independently developed, high-performance FTC862 processor cores. It is also compatible with the 64-bit ARMv8 instruction set and supports both ARM64 and ARM32 execution modes. With a clock speed of 2.5GHz, it supports single-precision and double-precision floating-point operations, ASIMD processing instructions, and hardware virtualization. It is suitable for both commercial and industrial applications.

Implications and Future Outlook

Phytium’s achievement represents a significant step forward for China’s efforts to develop a self-reliant AI ecosystem. By demonstrating comparable inference efficiency to foreign solutions, the company is positioning itself as a key player in the domestic AI infrastructure market.

The successful integration of Phytium CPUs with DeepSeek’s LLMs suggests a growing maturity in China’s domestic hardware and software capabilities. Further advancements in processor design, AI acceleration technology, and optimized software frameworks will be crucial to further close the gap with leading international competitors. This development could also spur innovation in AI applications tailored to the specific needs of the Chinese market.

References: