NVIDIA's Eos Supercomputer Breaks AI Training Speed Records

NVIDIA has a new supercomputer called Eos. It's super powerful and can train big AI models very quickly. This supercomputer uses over 10,000 special GPUs called H100 Tensor Core GPUs. With Eos, it can train a big AI model called GPT-3 with 175 billion parameters in less than four minutes. For a model called BERT, it only takes 7.2 seconds. Eos is incredibly fast because it has lots of GPUs working together. It has 10,752 of these GPUs, and they're connected to a super-fast network. Eos can process a huge amount of AI tasks, equivalent to 40 exaflops of power. You can rent access to Eos for about $37,000 a month. NVIDIA tested Eos with different AI models, and it set many records. For GPT-3, it only took 3.9 minutes. Other models, like Stable Diffusion, DLRM, RetinaNet, 3D U-Net, and BERT-Large, also performed exceptionally well. NVIDIA used a smaller version of GPT-3 with 175 billion parameters for the tests. The full-sized GPT-3 is much bigger and too challenging for benchmarking. The smaller version uses 1 billion tokens, which is a smaller unit of data that AI systems understand. The amazing improvement in performance was due to using more GPUs. They used 10,752 H100 GPUs in the recent tests, which is a big increase from the 3,584 GPUs used before. Even with more GPUs, they managed to make it work very efficiently through software improvements. Microsoft's Azure team also used a similar system with 10,752 H100 GPUs for testing and got results close to NVIDIA's. NVIDIA plans to use this supercomputer for plenty of AI tasks like growing AI models, GPU layout, neural rendering, generative AI, and self-sufficient driving. These benchmarks are crucial due to the fact they display how precise AI systems are. It's like a check to make certain AI systems are doing what they're speculated to do. NVIDIA thinks it's important because now and then human beings make large claims about AI, and it's difficult to know if the claims are real. NVIDIA has been focusing on AI a lot recently. They consider it's a good sized second for AI, like whilst the iPhone got here out. They're making powerful computer systems to help with AI duties and studies.