|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": 9.401168823242188, |
|
"base_token_generation_latency_sync": 35.41612300872803, |
|
"base_token_generation_latency_async": 35.34973859786987, |
|
"base_token_generation_throughput_sync": 0.02823572754571577, |
|
"base_token_generation_throughput_async": 0.02828875232645309, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 119.93333740234375, |
|
"base_inference_latency_async": 43.9906120300293, |
|
"base_inference_throughput_sync": 0.00833796525352473, |
|
"base_inference_throughput_async": 0.02273212292016693, |
|
"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": 9.79158878326416, |
|
"smashed_token_generation_latency_sync": 41.63551406860351, |
|
"smashed_token_generation_latency_async": 41.99160244315863, |
|
"smashed_token_generation_throughput_sync": 0.024017957322498378, |
|
"smashed_token_generation_throughput_async": 0.023814285281292532, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 162.9432815551758, |
|
"smashed_inference_latency_async": 70.10679244995117, |
|
"smashed_inference_throughput_sync": 0.006137104828476039, |
|
"smashed_inference_throughput_async": 0.0142639531071671, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |