SAGECal has been designed to find the most accurate calibration solutions for low radio frequency imaging observations, with minimum artefacts due to incomplete sky models. SAGECAL is developed to handle extremely large datasets, e.g., when the number of frequency bands greatly exceeds the number of available nodes on a compute cluster. Accurate calibration solutions are derived at the expense of large computational loads, which require distributed computing and modern compute devices, such as GPUs, to decrease runtimes. In this work, we investigate if the GPU version of SAGECal scales well enough to meet the requirements for the Square Kilometre Array and we compare its performance with the CPU version.