| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 31.86516227722168, | |
| "base_token_generation_latency_async": 32.29103069752455, | |
| "base_token_generation_throughput_sync": 0.03138223465803074, | |
| "base_token_generation_throughput_async": 0.030968351842564773, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 30.951731491088868, | |
| "base_inference_latency_async": 29.493379592895508, | |
| "base_inference_throughput_sync": 0.032308370221158844, | |
| "base_inference_throughput_async": 0.03390591426968526, | |
| "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": 63.732228469848636, | |
| "smashed_token_generation_latency_async": 64.06205799430609, | |
| "smashed_token_generation_throughput_sync": 0.01569064857779286, | |
| "smashed_token_generation_throughput_async": 0.015609863799394037, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 62.85414428710938, | |
| "smashed_inference_latency_async": 60.66994667053223, | |
| "smashed_inference_throughput_sync": 0.015909849880894614, | |
| "smashed_inference_throughput_async": 0.016482625333931706, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |