72 × NVIDIA GB200

Tested system

Processor: NVIDIA Grace CPU
Graphics: NVIDIA GB200
Accelerator count: 72
Nodes: 18

Recorded results

Only the same benchmark, version, workload, scenario and unit share a row. Missing results are not zero. System results include the full submitted hardware and software configuration.

MLPerf Inference / deepseek-r1Offline · v6.1 · Tokens/s

72 × NVIDIA GB200

510,781^* Tokens/s

Accuracy target / result: exact_match: 81.19872379216044 TOKENS_PER_SAMPLE: 3852.355970829535
Weight precision: fp4
Division: closed
System type: datacenter
Availability: available
Cores per processor: 144
Processors per node: 2
Nodes: 18
Total accelerators: 72
Software stack: PyTorch 2.10.0a0+b4e4ee81d3.nv25.12, TensorRT 10.14.1.48, CUDA Toolkit 13.1, cuDNN 9.17.0, TensorRT-LLM 1.3.0rc0
Operating system: Ubuntu 24.04

MLPerf Inference / deepseek-r1Server · v6.1 · Tokens/s

72 × NVIDIA GB200

388,590^* Tokens/s

Accuracy target / result: exact_match: 81.01640838650866 TOKENS_PER_SAMPLE: 3795.0934366453967
Weight precision: fp4
Division: closed
System type: datacenter
Availability: available
Cores per processor: 144
Processors per node: 2
Nodes: 18
Total accelerators: 72
Software stack: PyTorch 2.10.0a0+b4e4ee81d3.nv25.12, TensorRT 10.14.1.48, CUDA Toolkit 13.1, cuDNN 9.17.0, TensorRT-LLM 1.3.0rc0
Operating system: Ubuntu 24.04

Processing & platform

Processor: NVIDIA Grace CPU
Graphics: NVIDIA GB200

Connectivity & software

Operating system: Ubuntu 24.04

System configuration

Reported host core count: 144
Processors per node: 2
Accelerator count: 72
Nodes: 18
Software stack: PyTorch 2.10.0a0+b4e4ee81d3.nv25.12, TensorRT 10.14.1.48, CUDA Toolkit 13.1, cuDNN 9.17.0, TensorRT-LLM 1.3.0rc0
Submission division: closed

* Provisional data