| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 44.01970520019531, | |
| "base_token_generation_latency_async": 45.2003788203001, | |
| "base_token_generation_throughput_sync": 0.022717098977654285, | |
| "base_token_generation_throughput_async": 0.02212370838695021, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 42.85870094299317, | |
| "base_inference_latency_async": 41.524553298950195, | |
| "base_inference_throughput_sync": 0.023332485072986957, | |
| "base_inference_throughput_async": 0.024082137447708114, | |
| "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": 28.85710735321045, | |
| "smashed_token_generation_latency_async": 28.945353254675865, | |
| "smashed_token_generation_throughput_sync": 0.03465350798193384, | |
| "smashed_token_generation_throughput_async": 0.03454785958911933, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 30.32453098297119, | |
| "smashed_inference_latency_async": 28.869938850402832, | |
| "smashed_inference_throughput_sync": 0.032976602360694454, | |
| "smashed_inference_throughput_async": 0.03463810592678296, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |