NVIDIA / SemiAnalysis Misleading Marketing

NVIDIA and SemiAnalysis have been posting these ridiculous graphs, one of which was Jensen Huangs basis of comparing performance between Hopper and Blackwell and saying it’s 50x faster.

Sure. But they keep comparing NVL72, which is 72 GPUs versus 8 GPUs. Of course you’re going to get better per GPU perf when each GPU is under less stress.

In the graph above you can see that the B300 x 8 can reach the same throughput per GPU, albeit at a much lower tokens per second. So great, just buy 9 times the amount of GPUs for $5 million dollars.

At the actual speed providers serve on OpenRouter (30tps) it’s about a 2.5x improvement for 9x the amount of GPUs with a like for like product (B300). Congrats