Scrying the AMD GFX1250 LLVM Tea Leaves

# Summary AMD's upcoming MI455X accelerator (codenamed GFX1250) for machine learning workloads shows architectural similarities to consumer RDNA4 chips, including a redesigned workgroup processor structure with shared vector caching and support for Wave32-only operation, alongside new WMMA machine learning capabilities. The analysis of LLVM commits reveals that GFX1250 can run 20 waves per SIMD—four more than RDNA4—suggesting potential future directions for AMD's GPU architecture, though some expected features like dynamic VGPR allocation are notably absent.