LLM sources.twitter Mar 4, 2026 1 min read
NVIDIA AI Developer says a collaboration with SGLang achieved up to 25x faster DeepSeek R1 inference on GB300 NVL72 versus H200 and an 8x GB200 NVL72 gain within months. The post attributes gains to NVFP4 precision, disaggregation, and communication-compute overlap.