|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 11.945297241210938, |
|
"base_token_generation_latency_sync": 36.61009635925293, |
|
"base_token_generation_latency_async": 36.64432633668184, |
|
"base_token_generation_throughput_sync": 0.027314869378847112, |
|
"base_token_generation_throughput_async": 0.027289354177565443, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 120.30351409912109, |
|
"base_inference_latency_async": 37.97504901885986, |
|
"base_inference_throughput_sync": 0.008312309141494195, |
|
"base_inference_throughput_async": 0.026333079899471933, |
|
"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": 12.385092735290527, |
|
"smashed_token_generation_latency_sync": 43.115352630615234, |
|
"smashed_token_generation_latency_async": 42.25224871188402, |
|
"smashed_token_generation_throughput_sync": 0.023193594369211368, |
|
"smashed_token_generation_throughput_async": 0.023667379381839538, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 163.38503723144532, |
|
"smashed_inference_latency_async": 69.0446138381958, |
|
"smashed_inference_throughput_sync": 0.006120511504266062, |
|
"smashed_inference_throughput_async": 0.014483388991695618, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |