|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 9.401168823242188, |
|
"base_token_generation_latency_sync": 33.9238317489624, |
|
"base_token_generation_latency_async": 34.25745330750942, |
|
"base_token_generation_throughput_sync": 0.029477802136269177, |
|
"base_token_generation_throughput_async": 0.02919072795702518, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 121.07059249877929, |
|
"base_inference_latency_async": 44.10750865936279, |
|
"base_inference_throughput_sync": 0.00825964405856924, |
|
"base_inference_throughput_async": 0.022671876748307975, |
|
"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": 9.87337875366211, |
|
"smashed_token_generation_latency_sync": 165.7023483276367, |
|
"smashed_token_generation_latency_async": 165.62931071966887, |
|
"smashed_token_generation_throughput_sync": 0.006034917489658864, |
|
"smashed_token_generation_throughput_async": 0.006037578709075964, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 262.0567565917969, |
|
"smashed_inference_latency_async": 193.62993240356445, |
|
"smashed_inference_throughput_sync": 0.0038159672469643274, |
|
"smashed_inference_throughput_async": 0.005164490776745173, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |