ALERT🚨: GB300 NVL72 @sgl_project can get close to 1000 tok/s/user on DeepSeek Flash v4.1 by aggressively fusing kernels together!! TileRT started this trend in the ecosystem of megakernel engines on OSS frontier models! Glad to see SGLang widely adopting it! Great work by the SGLang, @radixark, and NVIDIA teams on pushing performance!

NVIDIA's Vera Rubin NVL72 could train large mixture-of-experts AI models using 75% fewer GPUs than Grace Blackwell. Training a 10T-parameter model on 100T tokens in one month would just require roughly 16,000 Rubin GPUs, compared with 64,000 GB200 GPUs.

BREAKING NEWS🚨: RUBIN NVL72 HITS 862 tok/s/user on GLM5.3 744B on agentic workloads! In less than 8 days, NVIDIA has increased interactivity on Rubin by close to 2x through low-latency specialized kernels.

Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar Jensen Sandbagging Performance Again, 2x more Annual Profit Per GigaWatt, The More you Buy, The More you Earn, AgentX, InferenceX, Extreme Co-Design

ALERT🚨 GB300 NVL72 NVLINK BACKPLANE ENABLES 13x BETTER PERF PER DOLLAR THAN HOPPER ON AGENTIC INFERENCE Through extreme co-design, GB300 NVL72 uses an innovative copper backplane to expand the scale-up domain to 72 versus 8 on Hopper. This allows optimizations like @vllm_project wide expert parallelism, which allows each GPU to have only a subset of experts.

NVIDIA GB300 NVL72 IS 7X BETTER 💰💰 PERF PER DOLLAR 💰💰 THAN H200 on long context agentic workloads due to disagg prefill and wide expert parallelism optimizations which take advantage of NVL72 copper backplane.

NVIDIA Vera Rubin NVL72 production racks are here. The compute tray is engineered for fast compute, assembly, and serviceability. Manufacturing is 100% automated, and every tray goes together in one minute. Congratulations to @Microsoft on the first operational Vera Rubin NVL72 racks, now rolling off @HonHai_Foxconn Ingrasys lines.

Kimi K3 decode on NVIDIA GB300 NVL72: 842 tokens/s at batch 1 with DSpark 🚀 A unified CuTe DSL decode path fuses attention, MoE and communication, delivering 1.5–2× higher decode throughput in several tested configurations and up to 1.8× higher P90 interactivity on the @SemiAnalysis_ AgentX workload. The post covers the kernel work, the correctness and long-context evals behind it, and benchmark sweeps across batch sizes, context lengths and parallelism configurations. Full write-up: Great work and collaboration with the @NVIDIA team!