Nvidia Blackwell Ultra Crushes MLPerf Inference v6.0: 2.49M Tokens/Sec on DeepSeek-R1, 30¢ per Million Tokens

NVIDIA Developer Blog·high signal

Nvidia set new MLPerf Inference v6.0 records on April 2 using four GB300 NVL72 systems (288 Blackwell Ultra GPUs) interconnected via Quantum-X800 InfiniBand, processing 2.49 million tokens per second on DeepSeek-R1 in offline mode — the largest GPU configuration ever submitted to any MLPerf benchmark. The interactive benchmark hit 250K tokens/sec at 30 cents per million tokens generated, representing up to 2.77x speedup over GB200 NVL72. Nvidia was the sole platform to submit across all new tests including Qwen3-VL-235B and text-to-video generation.