NVIDIA speeds JAX on Blackwell

NVIDIA showed it sped up training in JAX using Blackwell and NVFP4 (4-bit precision). All five sources are NVIDIA official—a single-vendor technical presentation, announcement-led but infrastructure-oriented with measured performance. The focus is combining low-precision arithmetic with the latest GPUs to raise large-scale training throughput; the through-line is less model novelty than pursuing training efficiency—the same model, faster and cheaper. It's a quiet but fundamental optimization meshing framework (JAX), precision format, and hardware. But the results are mainly in NVIDIA's own environment—reproducibility on other workloads and real-world cost impact remain to confirm.