| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 55.90721588134765, | |
| "base_token_generation_latency_async": 57.912600226700306, | |
| "base_token_generation_throughput_sync": 0.01788677873214628, | |
| "base_token_generation_throughput_async": 0.017267399427507577, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 57.48715438842773, | |
| "base_inference_latency_async": 57.23462104797363, | |
| "base_inference_throughput_sync": 0.01739519046712985, | |
| "base_inference_throughput_async": 0.017471942360932337, | |
| "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": 164.0231719970703, | |
| "smashed_token_generation_latency_async": 164.54514525830746, | |
| "smashed_token_generation_throughput_sync": 0.006096699556681305, | |
| "smashed_token_generation_throughput_async": 0.006077359489580642, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 170.87242279052734, | |
| "smashed_inference_latency_async": 128.0531644821167, | |
| "smashed_inference_throughput_sync": 0.005852319430303279, | |
| "smashed_inference_throughput_async": 0.007809256444729683, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |