| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 77.57367630004883, | |
| "base_token_generation_latency_async": 77.46337354183197, | |
| "base_token_generation_throughput_sync": 0.012890970851143876, | |
| "base_token_generation_throughput_async": 0.012909326747304356, | |
| "base_token_generation_CO2_emissions": 4.053060495387316e-05, | |
| "base_token_generation_energy_consumption": 0.013849622710902646, | |
| "base_inference_latency_sync": 75.81122741699218, | |
| "base_inference_latency_async": 74.6903657913208, | |
| "base_inference_throughput_sync": 0.013190658350637149, | |
| "base_inference_throughput_async": 0.013388607612311391, | |
| "base_inference_CO2_emissions": 4.126205610105293e-05, | |
| "base_inference_energy_consumption": 0.00011312198092062262, | |
| "smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "smashed_current_gpu_total_memory": 40339.3125, | |
| "smashed_token_generation_latency_sync": 49.07923049926758, | |
| "smashed_token_generation_latency_async": 48.95127713680267, | |
| "smashed_token_generation_throughput_sync": 0.020375217578338423, | |
| "smashed_token_generation_throughput_async": 0.020428476201046396, | |
| "smashed_token_generation_CO2_emissions": 3.559633476824905e-05, | |
| "smashed_token_generation_energy_consumption": 0.00914676739443409, | |
| "smashed_inference_latency_sync": 51.80733413696289, | |
| "smashed_inference_latency_async": 49.51908588409424, | |
| "smashed_inference_throughput_sync": 0.019302286378146828, | |
| "smashed_inference_throughput_async": 0.020194233842293213, | |
| "smashed_inference_CO2_emissions": 3.5450395516021004e-05, | |
| "smashed_inference_energy_consumption": 9.43633613009706e-05 | |
| } |