| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 76.22649383544922, | |
| "base_token_generation_latency_async": 76.26957688480616, | |
| "base_token_generation_throughput_sync": 0.013118798329603202, | |
| "base_token_generation_throughput_async": 0.013111387801591599, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 74.56542739868163, | |
| "base_inference_latency_async": 73.68848323822021, | |
| "base_inference_throughput_sync": 0.013411040946003892, | |
| "base_inference_throughput_async": 0.013570641653285207, | |
| "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": 164.62764434814454, | |
| "smashed_token_generation_latency_async": 165.28586857020855, | |
| "smashed_token_generation_throughput_sync": 0.006074313970533775, | |
| "smashed_token_generation_throughput_async": 0.006050123998200303, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 171.57990112304688, | |
| "smashed_inference_latency_async": 135.48994064331055, | |
| "smashed_inference_throughput_sync": 0.005828188461787605, | |
| "smashed_inference_throughput_async": 0.007380621729199734, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |