| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 49.34154014587402, | |
| "base_token_generation_latency_async": 49.31502193212509, | |
| "base_token_generation_throughput_sync": 0.020266898784342482, | |
| "base_token_generation_throughput_async": 0.020277796923143493, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 47.70785331726074, | |
| "base_inference_latency_async": 47.0245361328125, | |
| "base_inference_throughput_sync": 0.020960909587566773, | |
| "base_inference_throughput_async": 0.021265494191706147, | |
| "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": 183.24039611816406, | |
| "smashed_token_generation_latency_async": 183.97862035781145, | |
| "smashed_token_generation_throughput_sync": 0.005457311931126485, | |
| "smashed_token_generation_throughput_async": 0.005435414169620071, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 190.25786743164062, | |
| "smashed_inference_latency_async": 149.45664405822754, | |
| "smashed_inference_throughput_sync": 0.005256024434097573, | |
| "smashed_inference_throughput_async": 0.006690903614899884, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |