NVIDIA AI Updates: April 4, 2026

1. NVIDIA Sets New MLPerf Inference v6.0 Records with Blackwell Ultra

NVIDIA. NVIDIA’s Blackwell Ultra platform achieved the highest throughput and lowest token cost in the MLPerf Inference v6.0 benchmark round, with up to 2.77x token-per-second speedups over the GB200 NVL72 and 9x training performance wins over the nearest competitor. NVIDIA emphasized software-driven co-design improvements as a key factor, demonstrating that inference performance gains are increasingly coming from optimized software stacks rather than hardware alone. Source