| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 51.661082077026364, | |
| "base_token_generation_latency_async": 51.390972919762135, | |
| "base_token_generation_throughput_sync": 0.019356930977733024, | |
| "base_token_generation_throughput_async": 0.01945867033031895, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 50.84262351989746, | |
| "base_inference_latency_async": 50.39808750152588, | |
| "base_inference_throughput_sync": 0.019668536569688345, | |
| "base_inference_throughput_async": 0.019842022774569045, | |
| "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": 105.19943466186524, | |
| "smashed_token_generation_latency_async": 104.95399478822947, | |
| "smashed_token_generation_throughput_sync": 0.00950575450537568, | |
| "smashed_token_generation_throughput_async": 0.009527984161229368, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 108.70056991577148, | |
| "smashed_inference_latency_async": 113.0976676940918, | |
| "smashed_inference_throughput_sync": 0.009199583781160184, | |
| "smashed_inference_throughput_async": 0.008841915314335345, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |