|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 8.24533462524414, |
|
"base_token_generation_latency_sync": 36.94498176574707, |
|
"base_token_generation_latency_async": 37.84793596714735, |
|
"base_token_generation_throughput_sync": 0.02706727550552301, |
|
"base_token_generation_throughput_async": 0.02642152007623393, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 120.78776245117187, |
|
"base_inference_latency_async": 39.63320255279541, |
|
"base_inference_throughput_sync": 0.008278984391355434, |
|
"base_inference_throughput_async": 0.025231370053124005, |
|
"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": 8.612092018127441, |
|
"smashed_token_generation_latency_sync": 44.395752334594725, |
|
"smashed_token_generation_latency_async": 44.099863059818745, |
|
"smashed_token_generation_throughput_sync": 0.02252467741650962, |
|
"smashed_token_generation_throughput_async": 0.022675807374811157, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 164.30274658203126, |
|
"smashed_inference_latency_async": 71.32909297943115, |
|
"smashed_inference_throughput_sync": 0.006086325522870861, |
|
"smashed_inference_throughput_async": 0.014019524968421588, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |