| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 53.183304977416995, | |
| "base_token_generation_latency_async": 53.15421111881733, | |
| "base_token_generation_throughput_sync": 0.018802893133937914, | |
| "base_token_generation_throughput_async": 0.018813184862525145, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.04361000061035, | |
| "base_inference_latency_async": 51.90608501434326, | |
| "base_inference_throughput_sync": 0.018852412194201966, | |
| "base_inference_throughput_async": 0.019265563945415436, | |
| "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": 166.61248474121095, | |
| "smashed_token_generation_latency_async": 167.15945340692997, | |
| "smashed_token_generation_throughput_sync": 0.006001951183629722, | |
| "smashed_token_generation_throughput_async": 0.0059823119758929694, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 174.076416015625, | |
| "smashed_inference_latency_async": 141.44434928894043, | |
| "smashed_inference_throughput_sync": 0.00574460356485189, | |
| "smashed_inference_throughput_async": 0.0070699183461703004, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |