The Ask the Expert session explores running local multi-agent workflows on AMD GPUs, highlighting the advantages of local deployment for data security, performance, cost, and customization, supported by AMD’s powerful hardware and open-source software stack tailored for agentic AI. It also covers practical use cases, developer resources, and community support, emphasizing AMD’s commitment to enabling efficient, scalable, and collaborative AI development.
The Ask the Expert session on running local multi-agent workflows on AMD GPUs, hosted by Shailen Sobhee and Mahdi Ghodsi from AMD’s AI Developer Relations team, delves into the emerging field of agentic AI. They begin by explaining the foundational role of large language models (LLMs) as token processors and how agentic AI builds on this by incorporating planning, reasoning, memory, and tool usage to create autonomous agents capable of complex tasks. The discussion highlights the evolution from generative AI to AI agents and finally to agentic AI, which includes self-improving capabilities and goal-oriented workflows, emphasizing the dynamic and continuously evolving nature of AI technology.
The presenters emphasize the importance of running agentic AI locally, citing four main reasons: data sovereignty, performance, cost efficiency, and full customization. Local deployment ensures sensitive data remains secure, avoids cloud rate limits and latency issues, reduces operational costs by amortizing hardware investments, and allows users to fine-tune and optimize models without vendor restrictions. They showcase AMD’s powerful hardware lineup, including the MI300X and MI350X GPUs with large HBM memory and high bandwidth, designed to handle demanding AI workloads efficiently.
The AMD agentic AI software stack is presented from hardware to application layers, featuring the ROCm software stack for GPU management, inference engines like VLLM enhanced by AMD’s Atom plugin for optimized performance, and popular open-source agentic AI frameworks such as OpenClaw and Hermes Agent. These frameworks support multi-agent orchestration and tool integration, enabling sophisticated AI workflows. Real-world use cases demonstrate multi-agent systems handling diverse data types—such as images, PDFs, and emails—with specialized agents collaborating to produce structured outputs like reports or analytics, showcasing the flexibility and power of agentic AI on AMD platforms.
To facilitate adoption, AMD offers the Developer Program and Developer Cloud, providing access to high-end GPUs, cloud credits, expert support, and extensive training resources. Users can start locally on Ryzen AI machines or scale up to powerful MI300X-based nodes for production workloads. The program supports researchers, engineers, students, and open-source contributors, fostering a vibrant community with monthly events, office hours, and opportunities to win hardware through competitions. The open-source nature of AMD’s software stack ensures transparency and avoids vendor lock-in, encouraging innovation and collaboration.
The session concludes with a Q&A addressing practical concerns such as performance comparisons between different inference engines, availability of GPUs on cloud platforms, best practices for hosting large models, and legal considerations around deploying agentic AI in enterprise workflows. The experts advise tuning model serving parameters for efficiency, leveraging official Docker images, and consulting legal teams for compliance. They also encourage community engagement via AMD’s Discord and GitHub channels for technical support and updates, reinforcing AMD’s commitment to empowering developers in the agentic AI space.