AMD's VP of AI Software Anush Elangovan discusses the evolution of ROCm, AMD's open-source AI software stack, and its strategy to compete with Nvidia's CUDA ecosystem. Key developments include the unification of AMD's AI stacks under OneROCm, heavy investment in Triton as a GPU-agnostic programming layer that reduces the need for CUDA-to-HIP conversion, and a 100% open-source approach up to firmware. AMD is targeting a six-week release cadence, expanding ROCm support to consumer laptops with Strix Halo, and actively engaging the developer community by addressing GitHub complaints and direct X/Twitter outreach. Elangovan also notes that AI tools like Claude are now used to write and validate new AMD kernels, and the team is looking ahead to the MI450 GPU and differentiated ROCm features beyond CUDA parity.
Sort: