| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": 82.86221313476562, | |
| "base_token_generation_latency_sync": 20.888434410095215, | |
| "base_token_generation_latency_async": 20.86069267243147, | |
| "base_token_generation_throughput_sync": 0.04787338200495811, | |
| "base_token_generation_throughput_async": 0.04793704675595715, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 40.887705612182614, | |
| "base_inference_latency_async": 17.732644081115723, | |
| "base_inference_throughput_sync": 0.024457229502798192, | |
| "base_inference_throughput_async": 0.05639316931110935, | |
| "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": 92.91635131835938, | |
| "smashed_token_generation_latency_sync": 24.19665660858154, | |
| "smashed_token_generation_latency_async": 24.469297379255295, | |
| "smashed_token_generation_throughput_sync": 0.04132802379173914, | |
| "smashed_token_generation_throughput_async": 0.040867540432435345, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 52.49914817810058, | |
| "smashed_inference_latency_async": 26.189470291137695, | |
| "smashed_inference_throughput_sync": 0.019047928103662803, | |
| "smashed_inference_throughput_async": 0.038183284689740055, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |