|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_memory_inference_first": 690.0, |
|
"base_memory_inference": 570.0, |
|
"base_token_generation_latency_sync": 25.73595085144043, |
|
"base_token_generation_latency_async": 25.555139780044556, |
|
"base_token_generation_throughput_sync": 0.03885615129483473, |
|
"base_token_generation_throughput_async": 0.03913107142465634, |
|
"base_token_generation_CO2_emissions": 7.04025152217974e-06, |
|
"base_token_generation_energy_consumption": 0.00201063437593726, |
|
"smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"smashed_current_gpu_total_memory": 40339.3125, |
|
"smashed_memory_inference_first": 164.0, |
|
"smashed_memory_inference": 206.0, |
|
"smashed_token_generation_latency_sync": 20.38736572265625, |
|
"smashed_token_generation_latency_async": 21.09651416540146, |
|
"smashed_token_generation_throughput_sync": 0.04904998583945111, |
|
"smashed_token_generation_throughput_async": 0.047401195863912546, |
|
"smashed_token_generation_CO2_emissions": 6.962162215496025e-06, |
|
"smashed_token_generation_energy_consumption": 0.001568448312001272 |
|
} |