| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": 8.633415222167969, | |
| "base_token_generation_latency_sync": 31.724804496765138, | |
| "base_token_generation_latency_async": 32.258518412709236, | |
| "base_token_generation_throughput_sync": 0.031521076831284064, | |
| "base_token_generation_throughput_async": 0.030999563811523943, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 119.97163619995118, | |
| "base_inference_latency_async": 43.6870813369751, | |
| "base_inference_throughput_sync": 0.00833530350735024, | |
| "base_inference_throughput_async": 0.022890061990788975, | |
| "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": 9.325716972351074, | |
| "smashed_token_generation_latency_sync": 38.51363410949707, | |
| "smashed_token_generation_latency_async": 38.758694007992744, | |
| "smashed_token_generation_throughput_sync": 0.025964830977957754, | |
| "smashed_token_generation_throughput_async": 0.02580066293755362, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 161.8583541870117, | |
| "smashed_inference_latency_async": 69.56062316894531, | |
| "smashed_inference_throughput_sync": 0.00617824149406954, | |
| "smashed_inference_throughput_async": 0.01437594941568092, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |