| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 48.01503448486328, | |
| "base_token_generation_latency_async": 48.088258877396584, | |
| "base_token_generation_throughput_sync": 0.02082680999250869, | |
| "base_token_generation_throughput_async": 0.020795096835374095, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 47.12458267211914, | |
| "base_inference_latency_async": 45.2284574508667, | |
| "base_inference_throughput_sync": 0.02122034707358038, | |
| "base_inference_throughput_async": 0.0221099735954147, | |
| "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": 182.8625915527344, | |
| "smashed_token_generation_latency_async": 182.76504781097174, | |
| "smashed_token_generation_throughput_sync": 0.005468587049482001, | |
| "smashed_token_generation_throughput_async": 0.0054715056953026885, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 189.7058303833008, | |
| "smashed_inference_latency_async": 148.69062900543213, | |
| "smashed_inference_throughput_sync": 0.00527131927352733, | |
| "smashed_inference_throughput_async": 0.006725373392316922, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |