At CES 2026, AMD executives Andrej Zdravkovic, Chief Software Officer, and an unnamed product manager for the client side, discussed the evolution of ROCm and AMD’s broader strategy to democratize AI development across consumer and enterprise platforms. The conversation centered on how ROCm has transformed from a high-performance computing-focused stack in 2023 to a cross-platform, user-accessible framework today. Zdravkovic emphasized that ROCm 7.2 represents a significant leap forward, now supporting both Windows and Linux on Ryzen Max and newer Ryzen platforms. The goal is to make AI accessible to non-technical users—what AMD calls "practitioners"—by abstracting away the complexity of setup. This effort culminates in the upcoming "AI bundle" within the Adrenalin driver, which will allow users to optionally install a suite of AI tools—including PyTorch, ComfyUI, Amuse, LM Studio, and Ollama—without needing to navigate multiple configurations or GitHub repositories. The AI bundle is designed to reduce the intimidation factor for average users who may be curious about AI but overwhelmed by technical barriers. On Windows, the installation is optional during driver setup and managed through a new Download and Install Manager (DIME), enabling users to install, update, or remove components at any time. On Linux, especially with the new Halo box, ROCm will be pre-installed and optimized for immediate use, supporting developer workflows from day one. Despite progress, challenges remain. AMD acknowledged that while ROCm binaries are compiled from the same source for both Windows and Linux, differences in kernel-level architecture mean recompilation is still required for low-level development. Zdravkovic admitted that a universal intermediate format like NVIDIA’s PTX is not yet available, though AMD is evaluating options to improve portability across disparate hardware like Strix Halo (GFX1150), RDNA 4 (GFX1200), and MI350 (GFX950). The focus remains on optimizing libraries for consumer devices, recognizing that use cases differ vastly from those in large-scale data centers. Documentation was also identified as a key pain point. Zdravkovic noted that while the number of scattered documentation sources has reduced from 30 to around 15 in the past six months, AMD is actively working to consolidate resources. Feedback from developers, especially on the complexity of setup and lack of centralized guidance, is being taken seriously. On the topic of FSR 4, AMD confirmed that while the accidental early release was a misstep, the long-term plan remains open sourcing. The company is committed to sharing its technology through GPUOpen, though core components may be held back temporarily to protect competitive advantage, particularly in mobile and embedded markets where FSR has already been adopted by partners like Samsung and Arm. Regarding NPUs, AMD acknowledged the fragmented programming landscape. While Windows ML and Vitis are currently the primary access points, AMD is aligning with Microsoft’s vision of a unified abstraction layer. However, the lack of a standard programming model across different NPUs remains a challenge, with no immediate plans to integrate NPU programming into ROCm or OpenCL. On gaming and video encoding, AMD acknowledged past shortcomings in FFmpeg support for RDNA 4, particularly around AV1 encoding. While the team has not seen overwhelming demand, feedback from journalists and streamers suggests the issue persists. The company committed to reviewing concerns and improving support if feasible, emphasizing that user input directly influences the roadmap. Finally, on handhelds, AMD confirmed that its ADLX API allows OEMs like Asus and Lenovo to integrate Adrenalin features into their own software ecosystems. While Asus has done a strong job unifying features like RSR and AFMF, others have lagged. AMD stressed that it enables OEMs but cannot control their implementation choices. On SteamOS, collaboration is ongoing, though it remains a separate challenge due to the platform’s independent development. Overall, AMD is betting on simplicity, accessibility, and openness to bring AI to a broader audience—without forcing it on users who prefer gaming or traditional workflows. The message is clear: ROCm is no longer just for data centers. It’s becoming a bridge to AI for everyone.