|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 10.639251708984375, |
|
"base_token_generation_latency_sync": 37.03370628356934, |
|
"base_token_generation_latency_async": 37.292652763426304, |
|
"base_token_generation_throughput_sync": 0.027002428337659194, |
|
"base_token_generation_throughput_async": 0.02681493339569346, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 118.89151992797852, |
|
"base_inference_latency_async": 40.24190902709961, |
|
"base_inference_throughput_sync": 0.008411028815223952, |
|
"base_inference_throughput_async": 0.02484971573606467, |
|
"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": 12.093978881835938, |
|
"smashed_token_generation_latency_sync": 58.08485069274902, |
|
"smashed_token_generation_latency_async": 55.31206578016281, |
|
"smashed_token_generation_throughput_sync": 0.017216193001677704, |
|
"smashed_token_generation_throughput_async": 0.01807923797267831, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 193.5841293334961, |
|
"smashed_inference_latency_async": 110.74752807617188, |
|
"smashed_inference_throughput_sync": 0.005165712723677131, |
|
"smashed_inference_throughput_async": 0.009029546910629035, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |