12nm, about 5 years behind; it would be a very bad deal against current AMD Radeon RX offerings though the Chinese Economic Planning body would very likely throw a lot of support to secure domestic IP and production capacity.
FWIW, Pascal GPUs are completely usable for image classification, model training and even inferencing the full Stable Diffusion model. China doesn't need a dense node, just big server-grade GPU dies and the same software we have.