AMD is leaning into the growing buzz around AI agents with a new guide that walks users through running the OpenClaw AI agent on two of its newest hardware platforms: Ryzen AI MAX APUs and Radeon AI PRO GPUs. The goal is simple—make it easier for people to set up personal, fully local AI agents while showing just how much performance AMD’s latest chips and graphics cards can deliver for large language models.
To demonstrate how OpenClaw can be deployed across different types of systems, AMD outlines two main configurations. The first is RyzenClaw, designed around Ryzen AI MAX SoCs found in high-end laptops and mini PCs. The second is RadeonClaw, tailored for workstation-class setups powered by Radeon AI PRO graphics cards.
One of the biggest advantages AMD highlights with Ryzen AI MAX+ systems is memory capacity. These platforms support up to 128GB of fast unified memory, and AMD notes you can allocate as much as 112GB of that as VRAM to the integrated Radeon 8000S graphics. That’s a major deal for local AI because memory limits are often what keep large LLMs out of reach on compact computers. With this kind of headroom, AMD positions Ryzen AI MAX+ machines as capable options for running very large models locally, including configurations like Qwen 3.5 122B.
On performance, AMD shares figures for both single-agent and multi-agent workloads. For Qwen 3.5 122B A10B, Ryzen AI MAX+ is listed at up to 19 tokens per second on a single agent, along with multi-agent support up to two agents and 95K context concurrency. AMD also points out that multiple Ryzen AI MAX+ systems can be linked together to scale into faster workstation-style AI setups.
For more mainstream large-model workloads such as Qwen 3.5 35B A3B, Ryzen AI MAX+ is shown reaching around 45 tokens per second, and processing 10,000 input tokens in about 19.5 seconds. AMD also claims a maximum context window of 260K, and says multi-agent workflows can expand to a 6 x 95K concurrency range depending on the use case.
For users who want workstation GPU muscle, AMD’s RadeonClaw example focuses on the Radeon AI PRO R9700, described as its fastest 32GB RDNA 4-based GPU. AMD reports that a single AI PRO R9700 can handle 10,000 input tokens in just 4.4 seconds, delivering roughly 120 tokens per second. The context window is listed at 190K, with a multi-agent rate of 2 x 95K. For bigger local AI needs, AMD notes that workstation builds can scale up to four AI PRO R9700 GPUs, totaling 128GB of VRAM—enough to run larger 128B-class models locally with far fewer compromises.
To make setup less intimidating, AMD also shares a “Best Known Configuration” approach for getting OpenClaw running through WSL2. The configuration is positioned as a relatively quick start for early adopters building personal AI agents, and includes:
– Fully local LLM provisioning
– Functional Memory.md support for local embedding
– LM Studio (llama.cpp) as the engine underneath
– Browser control within WSL2
– Estimated setup time of under an hour
The broader message from AMD is clear: it’s not only shipping new AI-capable hardware, but also trying to lower the barrier for actually using that horsepower in real workflows. And while high-end Ryzen AI MAX+ laptops and mini PCs are still priced like compact workstations, AMD is clearly betting that interest in local AI agents—and the privacy, speed, and control they can offer—will keep growing across both professional and everyday PC users.
