| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 54.52168350219726, | |
| "base_token_generation_latency_async": 54.368472658097744, | |
| "base_token_generation_throughput_sync": 0.018341326528548942, | |
| "base_token_generation_throughput_async": 0.01839301255138456, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.9403263092041, | |
| "base_inference_latency_async": 53.14960479736328, | |
| "base_inference_throughput_sync": 0.018539005386575963, | |
| "base_inference_throughput_async": 0.01881481534646537, | |
| "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": 40.562165069580075, | |
| "smashed_token_generation_latency_async": 40.34195765852928, | |
| "smashed_token_generation_throughput_sync": 0.024653516356550654, | |
| "smashed_token_generation_throughput_async": 0.02478808808596762, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 51.46347541809082, | |
| "smashed_inference_latency_async": 41.28739833831787, | |
| "smashed_inference_throughput_sync": 0.01943125667040498, | |
| "smashed_inference_throughput_async": 0.02422046532953672, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |