Topic

Low-precision numerics

Use of reduced numeric formats such as FP8 and FP4 for weights, activations and KV cache to cut memory use and raise throughput, at some cost in accuracy.

Current clusters

build1 publisher

NVIDIA's 3.7x Vera Rubin figure comes from Qwen3-VL on vLLM and Dynamo

MLPerf Inference v6.1 preview submissions put Vera Rubin NVL72 at up to 3.7x GB300 on Qwen3-VL and up to 2.5x on DeepSeek-R1, on two different inference frameworks. The four-rack 99% scaling result is an offline number.

Publishers:nvidianews.nvidia.com

Reality

Evidence45
Adoption30
Hype gap+35
Incentives85
Confidence55