|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_token_generation_latency_sync": 76.91910247802734, |
|
"base_token_generation_latency_async": 76.82386375963688, |
|
"base_token_generation_throughput_sync": 0.013000671715919452, |
|
"base_token_generation_throughput_async": 0.013016788678173697, |
|
"base_token_generation_CO2_emissions": 3.9930267542867864e-05, |
|
"base_token_generation_energy_consumption": 0.013695786831045186, |
|
"base_inference_latency_sync": 75.31489334106445, |
|
"base_inference_latency_async": 74.6358871459961, |
|
"base_inference_throughput_sync": 0.013277586352960592, |
|
"base_inference_throughput_async": 0.013398380299866857, |
|
"base_inference_CO2_emissions": 4.072261623993283e-05, |
|
"base_inference_energy_consumption": 0.00011265253741674144, |
|
"smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"smashed_current_gpu_total_memory": 40339.3125, |
|
"smashed_token_generation_latency_sync": 49.13712501525879, |
|
"smashed_token_generation_latency_async": 48.9015968516469, |
|
"smashed_token_generation_throughput_sync": 0.02035121101793125, |
|
"smashed_token_generation_throughput_async": 0.02044922997164503, |
|
"smashed_token_generation_CO2_emissions": 3.524599859739772e-05, |
|
"smashed_token_generation_energy_consumption": 0.009180102018160505, |
|
"smashed_inference_latency_sync": 52.06343727111816, |
|
"smashed_inference_latency_async": 49.8340368270874, |
|
"smashed_inference_throughput_sync": 0.019207337287250974, |
|
"smashed_inference_throughput_async": 0.0200666063531993, |
|
"smashed_inference_CO2_emissions": 3.379301910600281e-05, |
|
"smashed_inference_energy_consumption": 9.774841757057516e-05 |
|
} |