| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 62.7690242767334, | |
| "base_token_generation_latency_async": 62.2707849368453, | |
| "base_token_generation_throughput_sync": 0.01593142495877015, | |
| "base_token_generation_throughput_async": 0.016058895050290352, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 60.96936950683594, | |
| "base_inference_latency_async": 60.41905879974365, | |
| "base_inference_throughput_sync": 0.01640167854922428, | |
| "base_inference_throughput_async": 0.016551068816124007, | |
| "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": 126.49608764648437, | |
| "smashed_token_generation_latency_async": 128.22878398001194, | |
| "smashed_token_generation_throughput_sync": 0.007905382835196265, | |
| "smashed_token_generation_throughput_async": 0.007798561048164335, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 131.06124725341797, | |
| "smashed_inference_latency_async": 129.16321754455566, | |
| "smashed_inference_throughput_sync": 0.007630020474827435, | |
| "smashed_inference_throughput_async": 0.007742142221372301, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |