| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": Infinity, | |
| "base_token_generation_latency_sync": 20.81943130493164, | |
| "base_token_generation_latency_async": 21.098139323294163, | |
| "base_token_generation_throughput_sync": 0.048032051661426656, | |
| "base_token_generation_throughput_async": 0.04739754462119387, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 39.95013084411621, | |
| "base_inference_latency_async": 18.52705478668213, | |
| "base_inference_throughput_sync": 0.02503120712925721, | |
| "base_inference_throughput_async": 0.053975119710815214, | |
| "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": 33.48402099609375, | |
| "smashed_token_generation_latency_async": 33.967031352221966, | |
| "smashed_token_generation_throughput_sync": 0.02986499142730379, | |
| "smashed_token_generation_throughput_async": 0.029440311978708868, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 65.11626243591309, | |
| "smashed_inference_latency_async": 40.59250354766846, | |
| "smashed_inference_throughput_sync": 0.015357146780102622, | |
| "smashed_inference_throughput_async": 0.024635090536499756, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |