| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 56.28690795898437, | |
| "base_token_generation_latency_async": 56.298492290079594, | |
| "base_token_generation_throughput_sync": 0.017766120688823207, | |
| "base_token_generation_throughput_async": 0.01776246502033254, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 55.02095375061035, | |
| "base_inference_latency_async": 53.02169322967529, | |
| "base_inference_throughput_sync": 0.01817489396008347, | |
| "base_inference_throughput_async": 0.018860204929108487, | |
| "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": 42.081617736816405, | |
| "smashed_token_generation_latency_async": 40.63092991709709, | |
| "smashed_token_generation_throughput_sync": 0.023763344989589576, | |
| "smashed_token_generation_throughput_async": 0.024611792101248708, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 51.81982650756836, | |
| "smashed_inference_latency_async": 40.34445285797119, | |
| "smashed_inference_throughput_sync": 0.019297633114498145, | |
| "smashed_inference_throughput_async": 0.024786555007212635, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |