Artificial Analysis published initial AA-AgentPerf results for NVIDIA’s Vera Rubin NVL72 on October 15, 2026, comparing its token throughput per megawatt of GPU power with GB300 NVL72 across DeepSeek V4 Pro, DeepSeek V4.1 Flash, GLM-5.2 and Kimi K3. At selected per-user output speeds, Vera Rubin’s throughput per megawatt is up to ~5.7x higher on DeepSeek V4 Pro, ~5.4x on DeepSeek V4.1 Flash, ~3.0x on GLM-5.2 and ~2.8x on Kimi K3. These are initial results, and the article says NVIDIA expects performance to improve as further optimizations arrive.

Vera Rubin NVL72 is a rack-scale system with 72 Rubin GPUs and 36 Vera CPUs in one NVLink domain, the same rack-scale count as GB300 NVL72. Each GPU has 288 GB HBM4 memory, 19.2 TB/s memory bandwidth, 3.0 TB/s NVLink bandwidth, 35 PFLOPS dense NVFP4 compute and 50 PFLOPS sparse NVFP4 compute. The listed Blackwell Ultra figures are 288 GB HBM3E, 8 TB/s, 1.8 TB/s, 15 PFLOPS dense and 20 PFLOPS sparse NVFP4. Vera has 88 CPU cores versus Grace’s 72, and its per-superchip CPU-to-GPU NVLink-C2C bandwidth is 1.8 TB/s versus 0.9 TB/s. Measured GPU power across the tested configurations ranges from 0.5 to 1.9 kW for Vera Rubin and 0.3 to 1.0 kW for GB300; those measurements exclude CPUs, networking and cooling.

At peak throughput per megawatt, Vera Rubin reaches 272M tokens/s on DeepSeek V4 Pro, 571M on DeepSeek V4.1 Flash, 272M on GLM-5.2 and 70M on Kimi K3. Those results are respectively 1.14x, 1.45x, 1.35x and 1.26x GB300’s peak figures. The biggest gaps emerge at higher per-user speeds: for example, at ~165 tok/s on DeepSeek V4 Pro, Vera Rubin delivers 174M tokens/s per MW versus 34M for GB300, a 5.1x difference; at ~190 tok/s, the reported gap reaches ~5.7x. From ~120 to ~165 tok/s per user, GB300 loses 81% of its throughput per MW, compared with 16% for Vera Rubin. At ~165 tok/s, the same GPU power budget serves ~5x as many DeepSeek V4 Pro tokens on Vera Rubin.

The fastest reported Vera Rubin configurations reach 906 tok/s per user on DeepSeek V4.1 Flash, 408 tok/s on GLM-5.2 and 255 tok/s on Kimi K3, each the highest result for that model on the leaderboard; 906 tok/s is the leaderboard’s fastest result across the four models. DeepSeek V4.1 Flash peaks at 571M tokens/s per MW, ahead of B300’s 433M and GB300’s 395M. The article notes a caveat: NVIDIA’s Vera Rubin Flash replies were capped at 2,048 output tokens, while other leaderboard results used 16,384; about 2% of replies reached the cap, and average replies were shorter, at 275 to 301 tokens versus 319 to 335. For DeepSeek V4 Pro, Vera Rubin’s median time to first token at ~165 tok/s was 1.65 seconds versus GB300’s 1.08 seconds; when that initial wait is included, GB300 matches or beats Vera Rubin’s throughput per MW at most speeds.

Artificial Analysis attributes most of the measured performance difference to upgraded rack hardware, alongside early Rubin-specific kernels and tuning, because the tested Vera Rubin and GB300 configurations are similar. NVIDIA’s configurations use KV-cache transfers between GPU and CPU memory, session-aware request routing through its open-source Dynamo serving framework, separate prefill and decode workers for many workloads, and speculative decoding with draft lengths adjusted to batch size. The configurations also use Rubin-tuned NVFP4 kernels and host-side request handling tuned for conversations of up to 132K tokens. Results are measured using total input and output tokens per second per megawatt of GPU power; about 96% of counted tokens are cached input. Per-user output speed is the 25th percentile across requests and excludes time to first token.

Artificial Analysis published initial AA-AgentPerf results for NVIDIA’s Vera Rubin NVL72 on October 15, 2026, comparing its token throughput per megawatt of GPU power with GB300 NVL72 across DeepSeek V4 Pro, DeepSeek V4.1 Flash, GLM-5.2 and Kimi K3. At selected per-user output speeds, Vera Rubin’s throughput per megawatt is up to ~5.7x higher on DeepSeek V4 Pro, ~5.4x on DeepSeek V4.1 Flash, ~3.0x on GLM-5.2 and ~2.8x on Kimi K3. These are initial results, and the article says NVIDIA expects performance to improve as further optimizations arrive.

Vera Rubin NVL72 is a rack-scale system with 72 Rubin GPUs and 36 Vera CPUs in one NVLink domain, the same rack-scale count as GB300 NVL72. Each GPU has 288 GB HBM4 memory, 19.2 TB/s memory bandwidth, 3.0 TB/s NVLink bandwidth, 35 PFLOPS dense NVFP4 compute and 50 PFLOPS sparse NVFP4 compute. The listed Blackwell Ultra figures are 288 GB HBM3E, 8 TB/s, 1.8 TB/s, 15 PFLOPS dense and 20 PFLOPS sparse NVFP4. Vera has 88 CPU cores versus Grace’s 72, and its per-superchip CPU-to-GPU NVLink-C2C bandwidth is 1.8 TB/s versus 0.9 TB/s. Measured GPU power across the tested configurations ranges from 0.5 to 1.9 kW for Vera Rubin and 0.3 to 1.0 kW for GB300; those measurements exclude CPUs, networking and cooling.

At peak throughput per megawatt, Vera Rubin reaches 272M tokens/s on DeepSeek V4 Pro, 571M on DeepSeek V4.1 Flash, 272M on GLM-5.2 and 70M on Kimi K3. Those results are respectively 1.14x, 1.45x, 1.35x and 1.26x GB300’s peak figures. The biggest gaps emerge at higher per-user speeds: for example, at ~165 tok/s on DeepSeek V4 Pro, Vera Rubin delivers 174M tokens/s per MW versus 34M for GB300, a 5.1x difference; at ~190 tok/s, the reported gap reaches ~5.7x. From ~120 to ~165 tok/s per user, GB300 loses 81% of its throughput per MW, compared with 16% for Vera Rubin. At ~165 tok/s, the same GPU power budget serves ~5x as many DeepSeek V4 Pro tokens on Vera Rubin.

The fastest reported Vera Rubin configurations reach 906 tok/s per user on DeepSeek V4.1 Flash, 408 tok/s on GLM-5.2 and 255 tok/s on Kimi K3, each the highest result for that model on the leaderboard; 906 tok/s is the leaderboard’s fastest result across the four models. DeepSeek V4.1 Flash peaks at 571M tokens/s per MW, ahead of B300’s 433M and GB300’s 395M. The article notes a caveat: NVIDIA’s Vera Rubin Flash replies were capped at 2,048 output tokens, while other leaderboard results used 16,384; about 2% of replies reached the cap, and average replies were shorter, at 275 to 301 tokens versus 319 to 335. For DeepSeek V4 Pro, Vera Rubin’s median time to first token at ~165 tok/s was 1.65 seconds versus GB300’s 1.08 seconds; when that initial wait is included, GB300 matches or beats Vera Rubin’s throughput per MW at most speeds.

Artificial Analysis attributes most of the measured performance difference to upgraded rack hardware, alongside early Rubin-specific kernels and tuning, because the tested Vera Rubin and GB300 configurations are similar. NVIDIA’s configurations use KV-cache transfers between GPU and CPU memory, session-aware request routing through its open-source Dynamo serving framework, separate prefill and decode workers for many workloads, and speculative decoding with draft lengths adjusted to batch size. The configurations also use Rubin-tuned NVFP4 kernels and host-side request handling tuned for conversations of up to 132K tokens. Results are measured using total input and output tokens per second per megawatt of GPU power; about 96% of counted tokens are cached input. Per-user output speed is the 25th percentile across requests and excludes time to first token.