|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_token_generation_latency_sync": 78.09549026489258, |
|
"base_token_generation_latency_async": 77.97934785485268, |
|
"base_token_generation_throughput_sync": 0.012804836701941352, |
|
"base_token_generation_throughput_async": 0.012823908220691406, |
|
"base_token_generation_CO2_emissions": 4.098068916180559e-05, |
|
"base_token_generation_energy_consumption": 0.013927917963664008, |
|
"base_inference_latency_sync": 76.36961288452149, |
|
"base_inference_latency_async": 75.5225419998169, |
|
"base_inference_throughput_sync": 0.013094213290200912, |
|
"base_inference_throughput_async": 0.013241079729578283, |
|
"base_inference_CO2_emissions": 4.097642058260856e-05, |
|
"base_inference_energy_consumption": 0.0001182669828520079, |
|
"smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"smashed_current_gpu_total_memory": 40339.3125, |
|
"smashed_token_generation_latency_sync": 49.69250984191895, |
|
"smashed_token_generation_latency_async": 49.56971164792776, |
|
"smashed_token_generation_throughput_sync": 0.02012375714531596, |
|
"smashed_token_generation_throughput_async": 0.020173609382732903, |
|
"smashed_token_generation_CO2_emissions": 3.5436386459368844e-05, |
|
"smashed_token_generation_energy_consumption": 0.009240227774657423, |
|
"smashed_inference_latency_sync": 52.84495315551758, |
|
"smashed_inference_latency_async": 49.048709869384766, |
|
"smashed_inference_throughput_sync": 0.018923282930294154, |
|
"smashed_inference_throughput_async": 0.020387896086624294, |
|
"smashed_inference_CO2_emissions": 3.4965807300444296e-05, |
|
"smashed_inference_energy_consumption": 0.00010078667242310871 |
|
} |