Scrying the AMD GFX1250 LLVM Tea Leaves
11 hours ago
- #GPU Architecture
- #AI Accelerators
- #AMD
- AMD's new GFX1250 accelerator, part of MI455X for machine learning, shows similarities to RDNA4 including WGP structure and WMMA support, but lacks dynamic VGPR allocation.
- GFX1250 allows up to 1024 VGPRs per wave, a major improvement over CDNA's 512 and RDNA's 256, and features a merged 448KB WGP Cache for LDS and vector L0.
- Graphics support is minimal in GFX1250, with no rasterizer, texture, or raytracing instructions, making it a pure compute accelerator.
- Tensor operations in GFX1250 combine RDNA4's programming model with CDNA4's performance, supporting various data formats and sparse WMMA, but excluding fp64.
- New features include clusters for thread block coordination, cluster-level barriers and loads, cooperative atomics, and hardware support for tanh and explicit prefetching.
- Data dependency handling is redesigned with more granular counters (e.g., ASYNCcnt, TENSORcnt) to reduce stalls, improving efficiency over GCN/CDNA.