| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 52.690575790405276, | |
| "base_token_generation_latency_async": 52.724254317581654, | |
| "base_token_generation_throughput_sync": 0.018978725986556703, | |
| "base_token_generation_throughput_async": 0.018966602997864224, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 52.11729965209961, | |
| "base_inference_latency_async": 52.47912406921387, | |
| "base_inference_throughput_sync": 0.019187486816763993, | |
| "base_inference_throughput_async": 0.01905519609437681, | |
| "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": 87.81638412475586, | |
| "smashed_token_generation_latency_async": 88.34282588213682, | |
| "smashed_token_generation_throughput_sync": 0.01138739666824992, | |
| "smashed_token_generation_throughput_async": 0.011319538287513654, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 92.48798599243165, | |
| "smashed_inference_latency_async": 85.88974475860596, | |
| "smashed_inference_throughput_sync": 0.01081221511388334, | |
| "smashed_inference_throughput_async": 0.01164283352815299, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |