|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 11.575923919677734, |
|
"base_token_generation_latency_sync": 40.255954360961915, |
|
"base_token_generation_latency_async": 38.7702826410532, |
|
"base_token_generation_throughput_sync": 0.024841045650870145, |
|
"base_token_generation_throughput_async": 0.025792950989248574, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 118.62886428833008, |
|
"base_inference_latency_async": 38.57831954956055, |
|
"base_inference_throughput_sync": 0.0084296516366327, |
|
"base_inference_throughput_async": 0.025921294957270663, |
|
"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": 12.261353492736816, |
|
"smashed_token_generation_latency_sync": 166.16905364990234, |
|
"smashed_token_generation_latency_async": 166.36428833007812, |
|
"smashed_token_generation_throughput_sync": 0.006017967714414962, |
|
"smashed_token_generation_throughput_async": 0.006010905405467378, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 263.9147018432617, |
|
"smashed_inference_latency_async": 195.67320346832275, |
|
"smashed_inference_throughput_sync": 0.0037891030435807153, |
|
"smashed_inference_throughput_async": 0.005110561805474241, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |