| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 55.36251564025879, | |
| "base_token_generation_latency_async": 55.878789350390434, | |
| "base_token_generation_throughput_sync": 0.018062763016368697, | |
| "base_token_generation_throughput_async": 0.017895878053646716, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.07125740051269, | |
| "base_inference_latency_async": 52.687764167785645, | |
| "base_inference_throughput_sync": 0.018842591055518113, | |
| "base_inference_throughput_async": 0.01897973876468685, | |
| "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": 164.15237579345703, | |
| "smashed_token_generation_latency_async": 165.58245047926903, | |
| "smashed_token_generation_throughput_sync": 0.00609190086446412, | |
| "smashed_token_generation_throughput_async": 0.006039287358687811, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 171.63048858642577, | |
| "smashed_inference_latency_async": 128.8050889968872, | |
| "smashed_inference_throughput_sync": 0.005826470624398664, | |
| "smashed_inference_throughput_async": 0.007763668406177388, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |