|
{ |
|
"base_current_gpu_type": "NVIDIA A100-PCIE-40GB", |
|
"base_current_gpu_total_memory": 40339.3125, |
|
"base_perplexity": Infinity, |
|
"base_token_generation_latency_sync": 31.82928295135498, |
|
"base_token_generation_latency_async": 32.45313633233309, |
|
"base_token_generation_throughput_sync": 0.0314176100519861, |
|
"base_token_generation_throughput_async": 0.03081366280779769, |
|
"base_token_generation_CO2_emissions": null, |
|
"base_token_generation_energy_consumption": null, |
|
"base_inference_latency_sync": 125.59482803344727, |
|
"base_inference_latency_async": 31.038379669189453, |
|
"base_inference_throughput_sync": 0.00796211130392797, |
|
"base_inference_throughput_async": 0.03221817667861895, |
|
"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": Infinity, |
|
"smashed_token_generation_latency_sync": 51.882550430297854, |
|
"smashed_token_generation_latency_async": 52.80750002712011, |
|
"smashed_token_generation_throughput_sync": 0.019274303049990966, |
|
"smashed_token_generation_throughput_async": 0.018936704056931958, |
|
"smashed_token_generation_CO2_emissions": null, |
|
"smashed_token_generation_energy_consumption": null, |
|
"smashed_inference_latency_sync": 211.0414840698242, |
|
"smashed_inference_latency_async": 107.99784660339355, |
|
"smashed_inference_throughput_sync": 0.004738404889482035, |
|
"smashed_inference_throughput_async": 0.009259443881991047, |
|
"smashed_inference_CO2_emissions": null, |
|
"smashed_inference_energy_consumption": null |
|
} |