|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 8.711935997009277, |
|
"base_token_generation_latency_sync": 37.35238075256348, |
|
"base_token_generation_latency_async": 36.53927445411682, |
|
"base_token_generation_throughput_sync": 0.02677205521715963, |
|
"base_token_generation_throughput_async": 0.027367812167582096, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 119.15110473632812, |
|
"base_inference_latency_async": 39.10989761352539, |
|
"base_inference_throughput_sync": 0.008392704391729478, |
|
"base_inference_throughput_async": 0.02556897514490474, |
|
"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": 9.512746810913086, |
|
"smashed_token_generation_latency_sync": 61.119361114501956, |
|
"smashed_token_generation_latency_async": 60.40619984269142, |
|
"smashed_token_generation_throughput_sync": 0.016361427570006574, |
|
"smashed_token_generation_throughput_async": 0.016554592121407724, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 194.49446563720704, |
|
"smashed_inference_latency_async": 109.34836864471436, |
|
"smashed_inference_throughput_sync": 0.0051415344736097145, |
|
"smashed_inference_throughput_async": 0.009145083848933467, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |