| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": 21.553375244140625, | |
| "base_token_generation_latency_sync": 21.522614669799804, | |
| "base_token_generation_latency_async": 21.35719507932663, | |
| "base_token_generation_throughput_sync": 0.04646275628412306, | |
| "base_token_generation_throughput_async": 0.046822627984888406, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 40.58705940246582, | |
| "base_inference_latency_async": 18.87080669403076, | |
| "base_inference_throughput_sync": 0.02463839496436261, | |
| "base_inference_throughput_async": 0.05299190523298197, | |
| "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": 23.815746307373047, | |
| "smashed_token_generation_latency_sync": 25.672776222229004, | |
| "smashed_token_generation_latency_async": 24.77431520819664, | |
| "smashed_token_generation_throughput_sync": 0.03895176709148195, | |
| "smashed_token_generation_throughput_async": 0.04036438511402921, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 51.47955169677734, | |
| "smashed_inference_latency_async": 26.007390022277832, | |
| "smashed_inference_throughput_sync": 0.019425188585366036, | |
| "smashed_inference_throughput_async": 0.0384506095822535, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |