| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 48.82857894897461, | |
| "base_token_generation_latency_async": 48.05881790816784, | |
| "base_token_generation_throughput_sync": 0.020479809601770108, | |
| "base_token_generation_throughput_async": 0.020807835971139128, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 46.71047592163086, | |
| "base_inference_latency_async": 44.975876808166504, | |
| "base_inference_throughput_sync": 0.021408473800989817, | |
| "base_inference_throughput_async": 0.022234141299017983, | |
| "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": 186.46332397460938, | |
| "smashed_token_generation_latency_async": 186.51706464588642, | |
| "smashed_token_generation_throughput_sync": 0.005362984948912364, | |
| "smashed_token_generation_throughput_async": 0.005361439726164245, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 192.6134780883789, | |
| "smashed_inference_latency_async": 164.21198844909668, | |
| "smashed_inference_throughput_sync": 0.005191744679160818, | |
| "smashed_inference_throughput_async": 0.0060896893670463375, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |