| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 53.581108474731444, | |
| "base_token_generation_latency_async": 53.849875181913376, | |
| "base_token_generation_throughput_sync": 0.018663294367483915, | |
| "base_token_generation_throughput_async": 0.01857014517901559, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 51.79688911437988, | |
| "base_inference_latency_async": 51.439499855041504, | |
| "base_inference_throughput_sync": 0.019306178751232753, | |
| "base_inference_throughput_async": 0.019440313432635206, | |
| "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": 40.933387756347656, | |
| "smashed_token_generation_latency_async": 41.39378983527422, | |
| "smashed_token_generation_throughput_sync": 0.024429934945829818, | |
| "smashed_token_generation_throughput_async": 0.024158213200083406, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 52.317081451416016, | |
| "smashed_inference_latency_async": 39.870476722717285, | |
| "smashed_inference_throughput_sync": 0.019114216088843655, | |
| "smashed_inference_throughput_async": 0.02508121502922043, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |