From Hopper to Blackwell, Nvidia got about 10x more ...

From Hopper to Blackwell, Nvidia got about 10x more performance and about 10x lower cost from only about 2x more hardware scale. Jensen’s explains Nvidia’s huge gain from Hopper to Blackwell did not come just from packing in more transistors. It came from redesigning the whole system together. --- Jensen Huang’s real claim is that Blackwell beats Hopper by shrinking communication, not merely by adding compute. That is a deeper point than it first sounds, because modern AI systems often stop being limited by arithmetic long before they stop being limited by movement. In his telling, the decisive bottleneck is mixture-of-experts inference, where each layer can require tokens to be routed all-to-all across many GPUs. Once that pattern dominates, a faster chip by itself does not rescue you. Amdahl’s law starts to bite, because the part of the workload spent waiting on communication, synchronization, and data movement becomes the ceiling on the whole machine. Huang argues that Nvidia’s advantage comes from co-designing the GPU, CPU, NIC, NVLink switches, system topology, and software stack so those all-to-all exchanges happen in one hop instead of being relayed chip to chip across a longer path. That sounds like plumbing until you notice what repeated multi-hop traffic does to an MoE model. If every layer is shuffling tokens and each shuffle crosses five or nine hops instead of one, latency compounds, bandwidth fragments, and expensive accelerators spend more time waiting than working. Now the 10x claim becomes less mystical. 10x higher performance and 10x lower cost from Hopper to Blackwell despite only about 2x more transistors, and his explanation is that system utilization improved far faster than raw silicon scale. Whether every benchmark comparison is settled is almost secondary. The interesting part is the mechanism: once AI becomes a network problem, the winning architecture is not the chip with the most flops, but the system that wastes the fewest of them. --- From 'NoRush Invest' YT channel