|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 6.637787342071533, |
|
"base_token_generation_latency_sync": 39.560679244995114, |
|
"base_token_generation_latency_async": 39.7673673927784, |
|
"base_token_generation_throughput_sync": 0.025277624628412102, |
|
"base_token_generation_throughput_async": 0.025146245918747846, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 118.31193618774414, |
|
"base_inference_latency_async": 38.94233703613281, |
|
"base_inference_throughput_sync": 0.008452232566062843, |
|
"base_inference_throughput_async": 0.02567899299603272, |
|
"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": 152463.125, |
|
"smashed_token_generation_latency_sync": 176.87936248779297, |
|
"smashed_token_generation_latency_async": 177.394749969244, |
|
"smashed_token_generation_throughput_sync": 0.0056535708063116365, |
|
"smashed_token_generation_throughput_async": 0.005637145406915233, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 270.4816131591797, |
|
"smashed_inference_latency_async": 232.70025253295898, |
|
"smashed_inference_throughput_sync": 0.0036971089765406545, |
|
"smashed_inference_throughput_async": 0.004297373935416606, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |