|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 8.405904769897461, |
|
"base_token_generation_latency_sync": 38.859296798706055, |
|
"base_token_generation_latency_async": 38.65551482886076, |
|
"base_token_generation_throughput_sync": 0.02573386762967101, |
|
"base_token_generation_throughput_async": 0.02586952998627212, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 118.89571762084961, |
|
"base_inference_latency_async": 39.14453983306885, |
|
"base_inference_throughput_sync": 0.00841073185822329, |
|
"base_inference_throughput_async": 0.025546347057967246, |
|
"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": 9.116766929626465, |
|
"smashed_token_generation_latency_sync": 165.78641815185546, |
|
"smashed_token_generation_latency_async": 166.55846014618874, |
|
"smashed_token_generation_throughput_sync": 0.006031857200051391, |
|
"smashed_token_generation_throughput_async": 0.006003897965448875, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 261.7901077270508, |
|
"smashed_inference_latency_async": 194.0150499343872, |
|
"smashed_inference_throughput_sync": 0.0038198540375812295, |
|
"smashed_inference_throughput_async": 0.005154239324929607, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |