Machine translation:
(Central News Agency, Taipei, 3rd) Chinese AI startup DeepSeek has reportedly developed a large language model that circumvents NVIDIA’s CUDA framework, as it prepares for future adaptation to domestically-produced Chinese GPU chips.
According to a report from Hong Kong's Sing Tao Daily, NVIDIA’s Compute Unified Device Architecture (CUDA) significantly reduces the difficulty of developing large AI models, making it widely adopted by developers worldwide and securing NVIDIA’s dominant position in artificial intelligence (AI) development.
The report, citing a U.S. tech website, states that while DeepSeek currently trains its models using NVIDIA’s H800 chips, it utilizes NVIDIA’s low-level hardware instruction language PTX (Parallel Thread Execution) instead of the higher-level programming language CUDA.
Huang Lei, an associate professor at Beihang University (Beijing University of Aeronautics and Astronautics), explained that bypassing CUDA means DeepSeek can develop directly based on GPU driver functions, enabling more fine-tuned operations.
The report further notes that DeepSeek has internal developers skilled in PTX language, which could facilitate its transition to Chinese domestic GPUs in the future. By understanding the basic function interfaces provided by GPU drivers, DeepSeek can mimic NVIDIA’s GPU hardware programming interfaces to develop relevant code. This capability would enhance its large language model’s adaptability to Chinese-made hardware.
(Edited by: Zhou Huiying / Zhang Shuling) 1140203
www.cna.com.tw
(Central News Agency, Taipei, 3rd) Chinese AI startup DeepSeek has reportedly developed a large language model that circumvents NVIDIA’s CUDA framework, as it prepares for future adaptation to domestically-produced Chinese GPU chips.
According to a report from Hong Kong's Sing Tao Daily, NVIDIA’s Compute Unified Device Architecture (CUDA) significantly reduces the difficulty of developing large AI models, making it widely adopted by developers worldwide and securing NVIDIA’s dominant position in artificial intelligence (AI) development.
The report, citing a U.S. tech website, states that while DeepSeek currently trains its models using NVIDIA’s H800 chips, it utilizes NVIDIA’s low-level hardware instruction language PTX (Parallel Thread Execution) instead of the higher-level programming language CUDA.
Huang Lei, an associate professor at Beihang University (Beijing University of Aeronautics and Astronautics), explained that bypassing CUDA means DeepSeek can develop directly based on GPU driver functions, enabling more fine-tuned operations.
The report further notes that DeepSeek has internal developers skilled in PTX language, which could facilitate its transition to Chinese domestic GPUs in the future. By understanding the basic function interfaces provided by GPU drivers, DeepSeek can mimic NVIDIA’s GPU hardware programming interfaces to develop relevant code. This capability would enhance its large language model’s adaptability to Chinese-made hardware.
(Edited by: Zhou Huiying / Zhang Shuling) 1140203

DeepSeekå³çºæ¹ç¨ä¸ååç¢GPUåæºå è·³åºè¼ééå¶ | ç§æ | ä¸å¤®ç¤¾ CNA
ä¸åAIæ°åµå¬å¸ã深度æ±ç´¢ãï¼DeepSeekï¼éç¼ç大èªè¨æ¨¡åç¹éç¾å人工æºæ§å·¨é è¼éï¼NVIDIAï¼çCUDAæ¡æ¶ï¼æå ±å°ï¼DeepSeekæ£çºæªä¾é©éä¸ååç¢GPUæ¶çåæºåã
