A little tour of the world of artificial intelligence, and Blackwell is back in top form! Indeed, early feedback suggests that it is up to 2.2 times better than Hopper at training AI. Clearly, the B200 is really getting the job done!
B200: 2.2x faster than Hopper!
To get an idea of the situation, you’ll need to look at the results of MLPerf Training v4.1 (an AI benchmark ), which shows that NVIDA’s new solution performs very well. The results show performance more than twice as good as that of the HGX H200 using the HGX B200 in LLM training. On GPT-3 pre-training, the scores are doubled while they are x2.2 Llama 2 70B.
Of course, in this kind of test, it’s not just one or two GPUs that are battling it out, but entire infrastructures. Here we learn that for the network part, ConnectX-7 network cards (€2,000 per card) and Quantum-2 switches are used for communication between the various nodes.
Next, we learn that whereas 256 GPUs were needed with Hopper to optimise performance on the GPT-3 175B, with Blackwell only 64 are required. In addition, it should be noted that the new generation of NVIDIA GPUs benefits from more memory with better bandwidth, higher-performance GPUs and improved NVLink bandwidth. All this combined means big performance gains.
