| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 56.57164268493652, | |
| "base_token_generation_latency_async": 57.186380587518215, | |
| "base_token_generation_throughput_sync": 0.017676700773376564, | |
| "base_token_generation_throughput_async": 0.017486681089557626, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 54.740274810791014, | |
| "base_inference_latency_async": 54.187893867492676, | |
| "base_inference_throughput_sync": 0.01826808512482785, | |
| "base_inference_throughput_async": 0.018454306462718974, | |
| "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": 172.90437774658204, | |
| "smashed_token_generation_latency_async": 173.57398755848408, | |
| "smashed_token_generation_throughput_sync": 0.00578354355761688, | |
| "smashed_token_generation_throughput_async": 0.0057612319338060936, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 180.40770568847657, | |
| "smashed_inference_latency_async": 157.52835273742676, | |
| "smashed_inference_throughput_sync": 0.005543000484285159, | |
| "smashed_inference_throughput_async": 0.006348063587428173, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |