| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 33.131116485595705, | |
| "base_token_generation_latency_async": 33.188691176474094, | |
| "base_token_generation_throughput_sync": 0.030183105976364137, | |
| "base_token_generation_throughput_async": 0.030130745279550314, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 32.26163177490234, | |
| "base_inference_latency_async": 30.899405479431152, | |
| "base_inference_throughput_sync": 0.030996572243377388, | |
| "base_inference_throughput_async": 0.03236308221741261, | |
| "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": 24.905581283569337, | |
| "smashed_token_generation_latency_async": 25.39039347320795, | |
| "smashed_token_generation_throughput_sync": 0.04015164266251108, | |
| "smashed_token_generation_throughput_async": 0.03938497452019419, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 26.685030364990233, | |
| "smashed_inference_latency_async": 24.64604377746582, | |
| "smashed_inference_throughput_sync": 0.037474193820366146, | |
| "smashed_inference_throughput_async": 0.0405744633511652, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |