| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 53.17919807434082, | |
| "base_token_generation_latency_async": 53.305395878851414, | |
| "base_token_generation_throughput_sync": 0.018804345236685772, | |
| "base_token_generation_throughput_async": 0.0187598269089442, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 52.69626922607422, | |
| "base_inference_latency_async": 51.61581039428711, | |
| "base_inference_throughput_sync": 0.01897667547791406, | |
| "base_inference_throughput_async": 0.01937390873767393, | |
| "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": 163.73534088134767, | |
| "smashed_token_generation_latency_async": 164.74530082195997, | |
| "smashed_token_generation_throughput_sync": 0.006107416973130189, | |
| "smashed_token_generation_throughput_async": 0.006069975865840924, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 171.19641571044923, | |
| "smashed_inference_latency_async": 139.2533779144287, | |
| "smashed_inference_throughput_sync": 0.005841243789188535, | |
| "smashed_inference_throughput_async": 0.007181154345961364, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |