Chinese chipmaker Sophgo has adapted its compute card to power DeepSeek’s reasoning model, demonstrating efforts by local firms to develop homegrown artificial intelligence infrastructure.
Sophgo’s SC11 FP300 compute card successfully passed verification, exhibiting stable and effective performance in executing reasoning tasks of DeepSeek’s R1 model. These tests were conducted by the China Telecommunication Technology Labs (CTTL), as confirmed by a company statement released on Monday. A compute card represents a compact module integrating a processor, memory, and other essential components required for computing tasks, commonly utilized in AI applications. CTTL operates as a research laboratory under the China Academy of Information and Communications Technology, an organization affiliated with the Ministry of Industry and Information Technology.
This verification signifies a milestone for Sophgo, confirming the compute card’s support for domestic AI models. This aligns with Beijing’s initiative to bolster its leading AI systems, such as the DeepSeek models, with self-developed infrastructure. Sophgo attributed this advancement to several innovations within the FP300 compute card, which was initially released last year. This card features 256 gigabytes of high-bandwidth memory and delivers up to 1.1 terabytes per second of memory bandwidth, facilitating faster data transfer during model training and execution. The card also enhances integration between hardware and algorithms to maximize performance specifically for reasoning models.
Researchers suspect DeepSeek cloned Gemini data
The announcement coincides with increased US chip export controls, which pose challenges to Chinese AI development. These controls have prompted domestic AI and chip companies to collaborate in navigating these restrictions. Separately, Chinese AI specialist iFlyTek confirmed it has been training and deploying its most recent AI models exclusively using chips from Huawei Technologies, thereby eliminating reliance on foreign semiconductor technology.
iFlyTek utilizes Huawei’s Ascend 910B chips for the development of its Xinghuo model series. iFlyTek chairman and president Liu Qingfeng stated in June that the company increased its training efficiency from 25 percent at the end of last year to 73 percent when benchmarked against Nvidia’s A800 chips. Liu acknowledged that using local chips generally resulted in a three-month delay in AI model development due to the requirement for additional computing resources compared to Nvidia chips and their established software ecosystem. Despite this, Liu affirmed the company’s commitment to improving training efficiency with local chips.
Both Sophgo and iFlyTek are subject to trade sanctions imposed by Washington. iFlyTek was added to the US Entity List in October 2019, which restricts its ability to acquire advanced US chips. Several Sophgo entities, including Sophgo Technologies, Sophon Technology (Beijing), and Xiamen Sophgo Technologies, were also added to this list earlier this year.
