| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 38.753916549682614, | |
| "base_token_generation_latency_async": 38.30525279045105, | |
| "base_token_generation_throughput_sync": 0.02580384356037918, | |
| "base_token_generation_throughput_async": 0.026106080162699923, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 32.50810871124268, | |
| "base_inference_latency_async": 32.19771385192871, | |
| "base_inference_throughput_sync": 0.030761555797743406, | |
| "base_inference_throughput_async": 0.03105810569653528, | |
| "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": 65.75935745239258, | |
| "smashed_token_generation_latency_async": 65.77239874750376, | |
| "smashed_token_generation_throughput_sync": 0.015206961240823623, | |
| "smashed_token_generation_throughput_async": 0.015203946017522323, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 70.72235565185547, | |
| "smashed_inference_latency_async": 67.43879318237305, | |
| "smashed_inference_throughput_sync": 0.014139800502724969, | |
| "smashed_inference_throughput_async": 0.014828260602109605, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |