apparently Intel 10nm second generation processing is doing better. Tiger Lake performance leaks are much improved over Ice Lake, with the Xe GPU in particular performing 2x better than the gen11 integrated GPU on Ice lake.
Intel is manufacturing small, scalable GPU chiplets, then stitching them with EMIB. They have some manufacturing expertise in this area from their FPGA production.
The Aurora project is at the end of next year, but note that Raja's site shows three other big discrete GPU chips are already in the lab.
While their Ponte Vecchio Xe is in Intel 7nm, I've seen rumors about them building some Xe GPUs at TSM.
Perhaps people don't know that they have been building 7nm eyeq5 chips at TSM since 2018.