|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 8.711685180664062, |
|
"base_token_generation_latency_sync": 39.507289505004884, |
|
"base_token_generation_latency_async": 39.948646537959576, |
|
"base_token_generation_throughput_sync": 0.02531178454733822, |
|
"base_token_generation_throughput_async": 0.025032137172652163, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 117.70030136108399, |
|
"base_inference_latency_async": 38.83824348449707, |
|
"base_inference_throughput_sync": 0.008496154966775952, |
|
"base_inference_throughput_async": 0.02574781736458207, |
|
"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": 9.484892845153809, |
|
"smashed_token_generation_latency_sync": 164.68993530273437, |
|
"smashed_token_generation_latency_async": 164.82422817498446, |
|
"smashed_token_generation_throughput_sync": 0.006072016472420077, |
|
"smashed_token_generation_throughput_async": 0.006067069211077131, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 259.89232330322267, |
|
"smashed_inference_latency_async": 191.85607433319092, |
|
"smashed_inference_throughput_sync": 0.0038477473566361398, |
|
"smashed_inference_throughput_async": 0.005212240495775644, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |