| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 82.928955078125, | |
| "base_token_generation_latency_async": 82.9247709363699, | |
| "base_token_generation_throughput_sync": 0.012058514412221022, | |
| "base_token_generation_throughput_async": 0.012059122849640709, | |
| "base_token_generation_CO2_emissions": 2.0673200656993337e-05, | |
| "base_token_generation_energy_consumption": 0.00675816312091373, | |
| "base_inference_latency_sync": 81.18558807373047, | |
| "base_inference_latency_async": 80.05075454711914, | |
| "base_inference_throughput_sync": 0.012317457121722492, | |
| "base_inference_throughput_async": 0.012492074630119622, | |
| "smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "smashed_current_gpu_total_memory": 40339.3125, | |
| "smashed_token_generation_latency_sync": 52.46407852172852, | |
| "smashed_token_generation_latency_async": 52.498188242316246, | |
| "smashed_token_generation_throughput_sync": 0.019060660706846116, | |
| "smashed_token_generation_throughput_async": 0.019048276397354767, | |
| "smashed_token_generation_CO2_emissions": 1.7540604124074345e-05, | |
| "smashed_token_generation_energy_consumption": 0.004500894375549395, | |
| "smashed_inference_latency_sync": 56.9942024230957, | |
| "smashed_inference_latency_async": 54.01430130004883, | |
| "smashed_inference_throughput_sync": 0.017545644249506526, | |
| "smashed_inference_throughput_async": 0.018513615393171735 | |
| } |