Switched to nvidia quadro k620. After installation of fitting drivers, matching requirements of cuda10 and escape from dependency hell deep learning on the GPU seems to work. Before: 24 min on CPU, now: 6 min on GPU – fair enough! (With a batch size of 512 (instead of default 128): 4.40 min, batch of 768: 4.21 min, batch of 1024: failed.