| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 66.22702407836914, | |
| "base_token_generation_latency_async": 66.2616390734911, | |
| "base_token_generation_throughput_sync": 0.015099576251782945, | |
| "base_token_generation_throughput_async": 0.015091688252548285, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 64.56258583068848, | |
| "base_inference_latency_async": 63.43851089477539, | |
| "base_inference_throughput_sync": 0.01548884678538806, | |
| "base_inference_throughput_async": 0.015763295605388448, | |
| "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": 178.2421890258789, | |
| "smashed_token_generation_latency_async": 179.08526603132486, | |
| "smashed_token_generation_throughput_sync": 0.00561034402385403, | |
| "smashed_token_generation_throughput_async": 0.005583932291923358, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 185.8609176635742, | |
| "smashed_inference_latency_async": 133.33563804626465, | |
| "smashed_inference_throughput_sync": 0.005380367279850056, | |
| "smashed_inference_throughput_async": 0.007499870362138449, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |