if i were to lean into cynicism, i might suggest this choice was meant to increase the effort required to reimplement cuda for other cards.