4 × NVIDIA RTX PRO 6000 Blackwell Server Edition

Tested system

Processor: AMD EPYC 9845
Graphics: NVIDIA RTX PRO 6000 Blackwell Server Edition
Accelerator count: 4
Nodes: 1

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 / llama3.1-8bOffline · v6.1 · Tokens/s

4 × NVIDIA RTX PRO 6000 Blackwell Server Edition

24,913.50^* Tokens/s

Accuracy target / result: ROUGE1: 38.7701 ROUGE2: 16.0155 ROUGEL: 24.534 ROUGELSUM: 35.6954 GEN_LEN: 8178898
Weight precision: fp4
Division: closed
System type: datacenter
Availability: available
Cores per processor: 160
Processors per node: 2
Nodes: 1
Total accelerators: 4
Software stack: TensorRT 10.14, CUDA 13.1
Operating system: Ubuntu 24.04

MLPerf Inference / llama3.1-8bServer · v6.1 · Tokens/s

4 × NVIDIA RTX PRO 6000 Blackwell Server Edition

24,498.60^* Tokens/s

Accuracy target / result: ROUGE1: 38.766 ROUGE2: 16.0082 ROUGEL: 24.5262 ROUGELSUM: 35.6898 GEN_LEN: 8179590
Weight precision: fp4
Division: closed
System type: datacenter
Availability: available
Cores per processor: 160
Processors per node: 2
Nodes: 1
Total accelerators: 4
Software stack: TensorRT 10.14, CUDA 13.1
Operating system: Ubuntu 24.04

MLPerf Inference / whisperOffline · v6.1 · Samples/s

4 × NVIDIA RTX PRO 6000 Blackwell Server Edition

9,236.94^* Samples/s

Accuracy target / result: ACCURACY: 97.8339693748394
Weight precision: fp16
Division: closed
System type: datacenter
Availability: available
Cores per processor: 160
Processors per node: 2
Nodes: 1
Total accelerators: 4
Software stack: TensorRT 10.14, CUDA 13.1
Operating system: Ubuntu 24.04

Processing & platform

Processor: AMD EPYC 9845
Graphics: NVIDIA RTX PRO 6000 Blackwell Server Edition

Connectivity & software

Operating system: Ubuntu 24.04

System configuration

Reported host core count: 160
Processors per node: 2
Accelerator count: 4
Nodes: 1
Software stack: TensorRT 10.14, CUDA 13.1
Submission division: closed

* Provisional data