Huawei Ascend 910C offers 60% of Nvidia H100 performance: Report

A new report cites that Huawei Ascend 910C chip offers 60% of Nvidia H100’s performance in DeepSeek AI models. The details arrive as the China-made AI LLMs are gaining popularity in the tech world for their cost-effective and efficient features.

DeepSeek R1 is the latest Chinese AI reasoning model. It performs excellently in tasks like maths, coding, and scientific problem-solving operations. The ultimate technology is already giving a tough spot to U.S. rivals like ChatGPT and Gemini.

While DeepSeek R1 has been trained on the Nvidia H100 processor, it uses Huawei Ascend 910C for inference. Now a new report adds more insight to this matter.

The DeepSeek team reportedly verified that Huawei Ascend 910C performance reaches 60% of the Nvidia H100 chips. The Chinese chipset performance is “unexpectedly good” in inference tasks. Besides, the handwritten CUNN kernel and other optimizations can further improve the Ascend chip’s performance.

If accurate, it is a great achievement for Huawei and the Chinese chip industry. Regardless of strict restrictions, the OEM managed to develop an improved AI chipset.

Other points in the report highlight that DeepSeek models supported Huawei AI chips since day one. The model converted CUDA (Compute Unified Device Architecture) to CUNN with one line of code. It obtained higher performance by just simple optimizations, which is another big achievement in the Chinese tech field.

Note: CUDA is a programming model and platform for parallel computing on Nvidia GPU processors. It allows developers to use the GPU for general purposes instead of only graphics processing.

DeepSeek and Huawei Ascend 910C chip integration show China is reducing its reliance on Nvidia and other foreign products. This measure eventually leads to substantial cost savings in the country in terms of technological products.

Some experts predicted:

• CUDA and PyTorch compilers’ importance will fade as AI models’ coverage increases in China.
• DeepSeek team will reduce reliance on Nvidia chips and save costs.
• Stability is still a challenge for China-made chips.

(source)