| Description: | 'Training a single AI model can emit as much carbon as five cars in their lifetimes'. Something similar could be said about calibration of radio interferometric data. The performance of SAGECal - Space Alternating Generalized Expectation Maximization Calibration - has been optimized for both CPUs and GPUs, see, e.g., https://arxiv.org/abs/1910.13908. We have also run SAGECal on the NVIDIA Jetson Nano low power GPU, see picture. This device has 128 cores and consumes 5-10W of power. We observed that its runtimes are comparable to SAGECal on a dual Intel Xeon E5-2660 v3 @ 2.60GHz (40 logical cores), while its power consumption is 20 times lower. This result, combined with lower hardware, maintenance and cooling costs make the case for GPU-based calibration for the SKA. |
