NVIDIA GB300 tops agentic inference benchmark at 261M tokens/s per MW
TheZachMueller · x · 2026-10-14
Artificial Analysis added DeepSeek V4.1 Flash to its AA-AgentPerf benchmark covering NVIDIA and AMD hardware: at 100 tokens/s per user, the NVIDIA GB300 ranks first with throughput up to 261M tokens/s per MW.
AA-AgentPerf replays real agentic coding sessions — including prompt caching and tool-call delays — to measure day-to-day agentic workloads realistically, reporting both total system throughput per megawatt and per-agent output speed at varying concurrency levels.
Sentdex added that the DGX Station runs on the GB300 and can be powered by a standard US home outlet: "It's kinda bananas."
More from Infra
- Applied Materials: yield becomes capacity, boosting the value of inspection and process control — BenBajarin · 2026-10-15
- MLPerf named the benchmark for EU's €10B AI Gigafactories tender as Endpoints v1.0 opens — TheKanter · 2026-10-15
- DecisionModels runs on Bittensor subnet 92, incentivizing community GPU hosting — airesearch12 · 2026-10-15
- DecisionModels launches Jev-compatible decision API: open models, 4x faster, from $0.019/1M tokens — airesearch12 · 2026-10-15
- Astera Labs and Lumentum demo PCIe over VCSEL optics with x8 PCIe 5 link-up — BenBajarin · 2026-10-15
- India's IndiaAI GPU Providers May Sell Compute Overseas as Domestic Demand Lags — deepakns · 2026-10-15