|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 7.471086502075195, |
|
"base_token_generation_latency_sync": 38.60911178588867, |
|
"base_token_generation_latency_async": 38.835852406919, |
|
"base_token_generation_throughput_sync": 0.02590062173783268, |
|
"base_token_generation_throughput_async": 0.025749402627296004, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 94.92889556884765, |
|
"base_inference_latency_async": 36.4933967590332, |
|
"base_inference_throughput_sync": 0.010534200298103594, |
|
"base_inference_throughput_async": 0.027402217628658264, |
|
"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_perplexity": 8.681138038635254, |
|
"smashed_token_generation_latency_sync": 53.12503852844238, |
|
"smashed_token_generation_latency_async": 51.72824040055275, |
|
"smashed_token_generation_throughput_sync": 0.01882351576017426, |
|
"smashed_token_generation_throughput_async": 0.019331800043005415, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 127.9923194885254, |
|
"smashed_inference_latency_async": 66.01173877716064, |
|
"smashed_inference_throughput_sync": 0.007812968809348366, |
|
"smashed_inference_throughput_async": 0.01514882077831268, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |