Deepseek states that Ascend 910C has 60% of the performance of a H100 in inference

DeepSeek has released Huawei Ai -facing AMA -facing AMA -facing, and concluded that the chips are able to achieve 60% of the performance of Nvidia’s GPU H100 in inference. Costing much less and with domestic manufacturing, this level of performance is already considered satisfactory to decrease China’s dependence on Nvidia GPUS.

As is well known, one of the pillars of Nvidia’s domain in the AI ​​segment is optimizing Hardware and AI software, through the CUDA platform. This is why DeepSek AI is optimized with native ascend accelerator support and a PyTorch repository that allows a simplified cuda conversion to CUNN – a system used by Huawei hardware.

The Ascend 910C is a new version of the Ascend 910, introduced in 2019 in China. The new accelerator maintains 53 billion transistors, but is all done at home. While the original version used TSMC lithography, based on its N7+technology, the 910C is manufactured by SMIC, with the manufacturer’s second generation 7nm architecture known as N+2.

Deepseek recognizes that AI training is still an obstacle in China

While DeepSeek was recently accused of distorting information to inflate the capabilities of its AI, the company does not avoid commenting current difficulties for the country’s technology in its recently disclosed tests.

It is important to note that the Ascend 910C would reach 60% of the H100 performance in inference. When we talk about AI training, the story is different. According to Deepseek Yuchen Jin, this remains the main weakness of Chinese hardware in the segment.

Related News:

NVIDIA dominates the world AI training market due to its integration between GPUS and CUDA. This is a challenge for company competitors around the world, but especially complicated for China who suffers US sanctions on importing these equipment.

Via: Tom’s Hardware