| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 56.49893798828125, | |
| "base_token_generation_latency_async": 56.557050719857216, | |
| "base_token_generation_throughput_sync": 0.01769944773488336, | |
| "base_token_generation_throughput_async": 0.017681261439060496, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 57.03167915344238, | |
| "base_inference_latency_async": 54.57870960235596, | |
| "base_inference_throughput_sync": 0.01753411463319401, | |
| "base_inference_throughput_async": 0.0183221627496454, | |
| "base_inference_CO2_emissions": null, | |
| "base_inference_energy_consumption": null, | |
| "smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "smashed_current_gpu_total_memory": 40339.3125, | |
| "smashed_token_generation_latency_sync": 166.02085876464844, | |
| "smashed_token_generation_latency_async": 166.74450617283583, | |
| "smashed_token_generation_throughput_sync": 0.006023339521557362, | |
| "smashed_token_generation_throughput_async": 0.0059971990859085285, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 173.71893768310548, | |
| "smashed_inference_latency_async": 141.13070964813232, | |
| "smashed_inference_throughput_sync": 0.0057564247936179504, | |
| "smashed_inference_throughput_async": 0.007085630069410153, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |