| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 30.974270248413085, | |
| "base_token_generation_latency_async": 30.937137454748154, | |
| "base_token_generation_throughput_sync": 0.03228486069179413, | |
| "base_token_generation_throughput_async": 0.03232361111181353, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 30.40849914550781, | |
| "base_inference_latency_async": 29.08501625061035, | |
| "base_inference_throughput_sync": 0.03288554279561436, | |
| "base_inference_throughput_async": 0.03438196463029361, | |
| "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": 88.86690368652344, | |
| "smashed_token_generation_latency_async": 88.7813925743103, | |
| "smashed_token_generation_throughput_sync": 0.011252783190550711, | |
| "smashed_token_generation_throughput_async": 0.011263621475220687, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 90.77841796875, | |
| "smashed_inference_latency_async": 86.16909980773926, | |
| "smashed_inference_throughput_sync": 0.01101583418587714, | |
| "smashed_inference_throughput_async": 0.011605088160735145, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |