| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 56.48856315612793, | |
| "base_token_generation_latency_async": 55.816091783344746, | |
| "base_token_generation_throughput_sync": 0.017702698460148727, | |
| "base_token_generation_throughput_async": 0.017915980285427208, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.93674201965332, | |
| "base_inference_latency_async": 53.28857898712158, | |
| "base_inference_throughput_sync": 0.018540237369836367, | |
| "base_inference_throughput_async": 0.018765747163978105, | |
| "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": 175.09743957519532, | |
| "smashed_token_generation_latency_async": 175.4741907119751, | |
| "smashed_token_generation_throughput_sync": 0.0057111057844483875, | |
| "smashed_token_generation_throughput_async": 0.0056988437783503385, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 180.2818572998047, | |
| "smashed_inference_latency_async": 157.45651721954346, | |
| "smashed_inference_throughput_sync": 0.005546869856887609, | |
| "smashed_inference_throughput_async": 0.006350959729445104, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |