Hardware portability in Triton comes from expressing kernels as tile operations rather than vendor-specific instructions. The compiler maps those operations to each target, allowing the same kernel source to run across different GPU vendors. section 1.1.1
A kernel written for NVIDIA GPUs can run on AMD GPUs through ROCm or Intel GPUs through the XPU backend without source changes. Toolchain updates can also expose newer hardware capabilities to existing kernels without a rewrite. section 1.1.1
Attention kernels across vendors
vLLM and SGLang run Triton attention kernels on AMD hardware through ROCm, avoiding separate vendor-specific implementations. section 1.1.1
