MosaicML claims that AMD AI chips are almost as fast as Nvidia’s.
Nvidia’s recent rise in market capitalization shows that the AI industry desperately needs new hardware for training large language models (LLMs), and other AI algorithms. Although server and HPC graphics cards are useless for gaming, these GPUs form the basis of data centers and supercomputers which perform highly parallelized calculations.
Nvidia GPUs are the best available for AI training. This is why, in recent weeks the company’s market capitalization briefly reached an unprecedented $1 trillion. MosaicML emphasizes now that Nvidia’s chips are just one option in a diverse hardware market. MosaicML suggests companies investing in AI shouldn’t blindly invest a fortune in Team Green.
The AI startup tested AMD MI250 cards and Nvidia A100 GPUs, which are both one generation behind the current flagship HPC graphics cards from each company. The AI startup used its own software tools for testing, as well as the open-source PyTorch software from Meta and AMD proprietary software.
Source:
https://www.techspot.com/news/99271-amd-ai-chips-punch-above-their-weight-mosaicml.html