| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 32.35535011291504, | |
| "base_token_generation_latency_async": 32.55457133054733, | |
| "base_token_generation_throughput_sync": 0.030906789650248217, | |
| "base_token_generation_throughput_async": 0.03071765221069453, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 31.760998153686522, | |
| "base_inference_latency_async": 30.65202236175537, | |
| "base_inference_throughput_sync": 0.03148515657981389, | |
| "base_inference_throughput_async": 0.03262427477697861, | |
| "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": 24.459324836730957, | |
| "smashed_token_generation_latency_async": 24.47169627994299, | |
| "smashed_token_generation_throughput_sync": 0.04088420292363443, | |
| "smashed_token_generation_throughput_async": 0.040863534287143, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 24.971366500854494, | |
| "smashed_inference_latency_async": 24.643588066101074, | |
| "smashed_inference_throughput_sync": 0.04004586613094566, | |
| "smashed_inference_throughput_async": 0.04057850655990991, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |