| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": Infinity, | |
| "base_token_generation_latency_sync": 34.424597930908206, | |
| "base_token_generation_latency_async": 35.069869086146355, | |
| "base_token_generation_throughput_sync": 0.02904899577932754, | |
| "base_token_generation_throughput_async": 0.0285145062145393, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 126.0896240234375, | |
| "base_inference_latency_async": 31.008625030517578, | |
| "base_inference_throughput_sync": 0.007930866696962474, | |
| "base_inference_throughput_async": 0.03224909195476535, | |
| "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_perplexity": Infinity, | |
| "smashed_token_generation_latency_sync": 41.684233856201175, | |
| "smashed_token_generation_latency_async": 41.889346577227116, | |
| "smashed_token_generation_throughput_sync": 0.023989885563201602, | |
| "smashed_token_generation_throughput_async": 0.023872418209159744, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 180.19839935302736, | |
| "smashed_inference_latency_async": 64.40317630767822, | |
| "smashed_inference_throughput_sync": 0.005549438860668769, | |
| "smashed_inference_throughput_async": 0.015527184485787214, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |