Over the past three days the new AMD “GFX1250” GPU target has started being built out within the upstream LLVM compiler codebase for the AMDGPU LLVM shader compiler back-end.
What AMD GFX1250 will correlate to remains to be seen. Based on prior AMD GFX1250 reference leaks there was some speculation that it could be for their upcoming UDNA architecture but that more likely will be for GFX13. AMD GFX1250 if following the convention used by RDNA3.5 with GFX1150 could end up being an RDNA4 refresh / “RDNA4.5” for potential future APUs. In fact, as part of the newly committed code to LLVM indicates that AMD GFX1250 is indeed for APUs as opposed to discrete GPUs:
Merged three days ago was stubbing out the new AMD GFX1250 target within LLVM. Following that were a number of GFX1250 commits foe new tests, renaming call instructions from b64 to i64, new s_setprio_inc_wg instruction, new s_wait_xcnt instruction, and the initial GFX1250 speed model.
It’s been just a few days with many more patches expected so over the coming weeks we’ll see more how this AMD GFX1250 compiler target is built-out and ultimately how it differs from the existing GFX1200 / RDNA4 support. As it stands right now GFX1250 seems to be an RDNA4 refresh for future APUs. In any event nice seeing the open-source driver support continuing to come early from AMD.
