8 × NVIDIA RTX PRO 4500 Blackwell

Tested system

Processor: AMD EPYC 9655
Graphics: NVIDIA RTX PRO 4500 Blackwell
Accelerator count: 8
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 / llama2-70b-99.9Offline · v6.1 · Tokens/s

8 × NVIDIA RTX PRO 4500 Blackwell

4,374.06^* Tokens/s

Accuracy target / result: ROUGE1: 44.7919 ROUGE2: 22.3659 ROUGEL: 29.1627 TOKENS_PER_SAMPLE: 273.2
Weight precision: fp4
Division: closed
System type: datacenter
Availability: available
Cores per processor: 96
Processors per node: 1
Nodes: 1
Total accelerators: 8
Software stack: TensorRT 10.14, CUDA 13.3
Operating system: Ubuntu 24.04.3

MLPerf Inference / llama2-70b-99.9Server · v6.1 · Tokens/s

8 × NVIDIA RTX PRO 4500 Blackwell

3,012.33^* Tokens/s

Accuracy target / result: ROUGE1: 44.7861 ROUGE2: 22.3547 ROUGEL: 29.1687 TOKENS_PER_SAMPLE: 271.3
Weight precision: fp4
Division: closed
System type: datacenter
Availability: available
Cores per processor: 96
Processors per node: 1
Nodes: 1
Total accelerators: 8
Software stack: TensorRT 10.14, CUDA 13.3
Operating system: Ubuntu 24.04.3

Processing & platform

Processor: AMD EPYC 9655
Graphics: NVIDIA RTX PRO 4500 Blackwell

Connectivity & software

Operating system: Ubuntu 24.04.3

System configuration

Reported host core count: 96
Processors per node: 1
Accelerator count: 8
Nodes: 1
Software stack: TensorRT 10.14, CUDA 13.3
Submission division: closed

* Provisional data