|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 9.883513450622559, |
|
"base_token_generation_latency_sync": 37.07163429260254, |
|
"base_token_generation_latency_async": 37.12908215820789, |
|
"base_token_generation_throughput_sync": 0.02697480213866765, |
|
"base_token_generation_throughput_async": 0.026933065453624103, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 121.7193977355957, |
|
"base_inference_latency_async": 105.3497314453125, |
|
"base_inference_throughput_sync": 0.00821561738394602, |
|
"base_inference_throughput_async": 0.009492193157788013, |
|
"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": 144668.515625, |
|
"smashed_token_generation_latency_sync": 175.58663482666014, |
|
"smashed_token_generation_latency_async": 175.85825938731432, |
|
"smashed_token_generation_throughput_sync": 0.005695194289629186, |
|
"smashed_token_generation_throughput_async": 0.005686397690298849, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 268.8863250732422, |
|
"smashed_inference_latency_async": 232.02152252197266, |
|
"smashed_inference_throughput_sync": 0.003719043724992742, |
|
"smashed_inference_throughput_async": 0.0043099449961815466, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |