Vera Rubin NVL72 posts first AgentPerf results
Artificial Analysis's first independent agentic-coding benchmarks of NVIDIA's Vera Rubin NVL72 show up to ~5.7x GB300's throughput per megawatt at high per-user speeds, but only 1.14–1.45x at peak throughput.
Artificial Analysis has published the first AA-AgentPerf results for NVIDIA's Vera Rubin NVL72, the rack-scale successor to Blackwell with 72 Rubin GPUs and 36 Vera CPUs. The results are the first independent figures on how much the Blackwell-to-Rubin move improves serving efficiency.
The benchmark replays real coding-agent trajectories at a range of concurrency levels. Requests carry up to 132K input tokens, and on average 97% of each request's input repeats the previous one, so KV cache capacity and reuse matter more than in chat-style tests.
The headline gains depend on where you measure. Vera Rubin's largest leads in throughput per MW over GB300 NVL72 come at high per-user output speeds, where GB300's throughput falls off:
- DeepSeek V4 Pro: ~5.7x at ~190 tok/s per user
- DeepSeek V4.1 Flash: ~5.4x at ~580 tok/s
- GLM-5.2: ~3.0x at ~165 tok/s
- Kimi K3: ~2.8x at ~67 tok/s
At peak throughput the gap is much smaller. On DeepSeek V4 Pro, Vera Rubin serves 272M tokens/s per MW against 239M for GB300 (1.14x). The other peak figures are 1.45x on V4.1 Flash, 1.35x on GLM-5.2 and 1.26x on Kimi K3. On DeepSeek V4 Pro, going from ~120 to ~165 tok/s per user costs GB300 81% of its throughput per MW, against 16% for Vera Rubin.
Per GPU, Rubin leads by 2.0–2.5x at peak throughput, and by up to ~12x on DeepSeek V4 Pro at ~165 tok/s. Each Rubin GPU has the same 288 GB of memory as Blackwell Ultra but 2.4x the memory bandwidth, 1.7x the NVLink bandwidth and 2.3x the dense NVFP4 compute. Rubin GPUs drew 0.5–1.9 kW in testing, against 0.3–1.0 kW for GB300.
Power figures count GPU and HBM draw only. Per-user speed is the 25th-percentile speed and leaves out time to first token.