HipKittens
lastmod 2025-11-17created 2025-11-171 edits
Fast and powerful AMD kernels
About
- HipKittens is a project that improves AI compute efficiency by providing high-performance kernels and C++-based programming primitives for AMD GPUs
- Existing tools in the AMD ecosystem, such as AITER, PyTorch, Triton, TileLang, and Composable Kernel, have shown limits due to unstable performance and immature support
- HipKittens centers on a tile abstraction, keeping a common interface between NVIDIA and AMD while separating hardware-specific implementations
- Kernels written in roughly 500 lines of code or fewer achieve faster performance than AMD's existing hand-written assembly kernels
- Presents a practical foundation for scaling AI compute to multi-silicon environments, and suggests the possibility of a shift toward an open hardware ecosystem