AMD recently announced the adoption of SPIR-V as a portable intermediate representation (IR) within its ROCm platform. This marks a shift from ROCm's prior ahead-of-time (AOT) compilation model, now enabling what AMD describes as a "compile-once, specialize-on-device model" for GPU code. This approach mirrors the just-in-time (JIT) compilation widely adopted in modern CPU software, addressing the increasing diversity of target architectures. By integrating SPIR-V, ROCm aims to offer greater flexibility and portability for developers. Code can now be compiled once to SPIR-V and then specialized at runtime for various AMD GPUs, including MI350X (CDNA4) and RX 9070 XT (RDNA4). AMD states this provides a more efficient and scalable deployment strategy, reducing the need to enumerate specific hardware at build time.