|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_token_generation_latency_sync": 56.065922546386716, |
|
"base_token_generation_latency_async": 56.165905855596066, |
|
"base_token_generation_throughput_sync": 0.017836146353832665, |
|
"base_token_generation_throughput_async": 0.017804395473848936, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 57.000243759155275, |
|
"base_inference_latency_async": 55.65207004547119, |
|
"base_inference_throughput_sync": 0.017543784623541753, |
|
"base_inference_throughput_async": 0.01796878353640643, |
|
"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_token_generation_latency_sync": 173.6800567626953, |
|
"smashed_token_generation_latency_async": 174.10976719111204, |
|
"smashed_token_generation_throughput_sync": 0.005757713456797935, |
|
"smashed_token_generation_throughput_async": 0.005743503171205481, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 179.55246124267578, |
|
"smashed_inference_latency_async": 157.53896236419678, |
|
"smashed_inference_throughput_sync": 0.0055694029092056875, |
|
"smashed_inference_throughput_async": 0.006347636070423083, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |