AMD MI250 graphics card catches up with NVIDIA A100 in the AI ​​field, with performance reaching 80%!

July 3 news, it is understood that in the field of graphics cards, NVIDIA has always been recognized as a strong competitor. It has a leading edge in both gaming and computing, and it is difficult to be caught up by other competitors. Especially in the field of AI, NVIDIA almost monopolizes the market. However, the good news is that AMD has finally begun to exert its strength. The performance of their MI250 graphics card has reached that of NVIDIA. 80% level of A100 graphics card.

#AMD’s lagging in the AI ​​field in the past was mainly due to the lagging of its software ecosystem. Although AMD's graphics card hardware specifications are excellent, there is a considerable gap between its computing environment and NVIDIA's CUDA. AMD recently upgraded the MI250 accelerator card to better support the PyTorch framework. According to research from MosaicML, after this wave of optimization, the MI250 graphics card has reached 80% of the performance level of the A100 graphics card in terms of large language model training speed.

It is worth noting that AMD emphasized that they did not fund MosaicML to conduct this research, but they promised to continue working with the startup to further optimize software support. However, it should be pointed out that the A100 graphics card was released in March 2020 and belongs to the previous generation of products, while NVIDIA's current top AI accelerator card is the H100, whose AI performance has been improved by several to dozens of times. Similarly, AMD's MI250 graphics card is not the latest product. It was released at the end of 2021. It uses the CDNA2 architecture and 6nm process. It has 208 computing units and 13312 stream processor cores. Its performance is about 5.5% lower than the MI250X. But other specifications are exactly the same.

The above is the detailed content of AMD MI250 graphics card catches up with NVIDIA A100 in the AI ​​field, with performance reaching 80%!. For more information, please follow other related articles on the PHP Chinese website!