We still don't have much confirmed information on AMD's GFX1250 that has come to light in recent months due to being developed for their AMDGPU shader compiler back-end within LLVM. AMD GFX1250 surfaced in LLVM activity and the past few months has been seeing more additions made. A notable change committed today for LLVM is that the number of user Scalar General-Purpose Registers (SGPRs) has doubled.

AMD GFX1250 from the code is noted as being for an "APU" product and being part of the "GFX12" family puts it as an RDNA4 target. AMD RDNA 3.5 as the refresh product for APUs ended up being GFX1150. But there's also been some code snippets implying GFX1250 as being CDNA architecture even with being in the GFX12 series rather than GFX9 derived. So whether this ends up being some APU compute accelerator or a more general purpose product remains to be seen. But in any event the LLVM activity by AMD for their AMDGPU GFX1250 target continues.

An interesting pull request was merged today. The patch confirms that the user SGPR count increases from 16 with prior GPUs to now 32 with GFX1250. SGPRs are used for data that has the same values for all threads in a wave.

This doubling of the user SGPRs can be very helpful for performance in reducing the chances of register spilling. Today's LLVM commit didn't provide any other insight into this change beyond that GFX1250 will double the user SGPR count from 16 to 32.

The continued LLVM activity with the refinements compared to GFX1200 are making me very eager to learn more about the GFX1250 hardware.