|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_token_generation_latency_sync": 57.0241153717041, |
|
"base_token_generation_latency_async": 57.01630376279354, |
|
"base_token_generation_throughput_sync": 0.017536440389853192, |
|
"base_token_generation_throughput_async": 0.01753884299761568, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 53.97422103881836, |
|
"base_inference_latency_async": 52.460384368896484, |
|
"base_inference_throughput_sync": 0.01852736326256192, |
|
"base_inference_throughput_async": 0.019062002919538183, |
|
"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_token_generation_latency_sync": 174.07499084472656, |
|
"smashed_token_generation_latency_async": 175.05259327590466, |
|
"smashed_token_generation_throughput_sync": 0.005744650596546586, |
|
"smashed_token_generation_throughput_async": 0.005712568898787324, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 179.4249710083008, |
|
"smashed_inference_latency_async": 156.67836666107178, |
|
"smashed_inference_throughput_sync": 0.005573360242893598, |
|
"smashed_inference_throughput_async": 0.006382502072945463, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |