Blackwell completes the transition from Hopper to the next era of AI compute. Highlights from our early production benches:
- 4× compute over H100 for FP8 workloads.
- 3× energy efficiency per operation — important for power-constrained facilities.
- 800 Gb/s InfiniBand support on the reference NIC, unlocking larger pre-training runs.
The platform changes the economics of 1T+ parameter training, especially when combined with FP4 numerics and MegaScale optimizations. Expect the industry to re-center around Blackwell by late 2026 as supply ramps.