AMD has entered the local AI agent race with a clear hardware pitch. The company has published an official guide for running OpenClaw locally on AMD hardware via two configurations — RyzenClaw (Ryzen AI Max+ with 128GB unified memory) and RadeonClaw (Radeon AI PRO R9700 GPU) — positioning both as serious alternatives to the Mac mini-dominated OpenClaw setup that’s been making waves. The catch: you’ll need a very specific (and expensive) AMD system to do it properly.
RyzenClaw vs RadeonClaw: Performance Compared
| Metric | RyzenClaw (Ryzen AI Max+ 128GB) | RadeonClaw (Radeon AI PRO R9700) |
|---|---|---|
| Model Used | Qwen 3.5 35B A3B | Qwen 3.5 35B A3B |
| Tokens Per Second | ~45 tokens/s | ~120 tokens/s |
| 10K Input Tokens | ~19.5 seconds | ~4.4 seconds |
| Max Context Window | 260K tokens | 190K tokens |
| Concurrent Agents | Up to 6 | Up to 2 (×95K) |
| VRAM Allocation | Up to 96GB reserved from unified pool | 32GB GDDR6 (up to 128GB with 4× GPUs) |
| Setup Path | WSL2 + LM Studio on Windows | Same |
| Entry Cost | ~$2,000+ (Mini PC, 64GB); ~$4,000+ (128GB laptop) | GPU card only (professional card, ~$2,500+) |
What Is RyzenClaw — and Why Does It Need 128GB?
RyzenClaw isn’t just any Ryzen AI Max+ machine — AMD’s own guide specifically calls for the 128GB unified memory configuration. Here’s why that matters: large language models like Qwen 3.5 35B need their parameters loaded into memory to run. On a system with 128GB, AMD recommends reserving 96GB as variable graphics memory — giving the GPU tile more than enough headroom to handle serious AI workloads without swapping to disk, which kills performance entirely.
This unified memory architecture is AMD’s answer to Apple‘s approach on the Mac mini and Mac Studio. Where Apple uses M-series chips with shared CPU-GPU memory, AMD’s Ryzen AI Max+ does the same on x86 — and at 128GB, it can run models that would require a $10,000+ NVIDIA workstation otherwise. With the Qwen 3.5 35B model, the RyzenClaw delivers around 45 tokens per second with a 260K token context window and can run up to 6 concurrent agents — enough for real multi-agent workflows.
The practical form factors for RyzenClaw are mini PCs like the GMKtec EVO X2 and laptops like the HP ZBook Ultra G1a. Both are compact, both are genuinely capable — and both are expensive entry points for what is still early-adopter territory.
RadeonClaw Is Faster But Less Practical
The Radeon AI PRO R9700 — AMD’s 32GB RDNA 4 professional GPU — is the faster of the two configurations. With the same Qwen 3.5 35B model, the RadeonClaw achieves roughly 120 tokens per second and processes 10,000 input tokens in about 4.4 seconds. That’s nearly 3× the speed of RyzenClaw at the same model size.
The trade-off: it’s a professional workstation card, not a consumer GPU. Four of them can be stacked for 128GB of discrete VRAM to run 128B+ models — but that configuration costs tens of thousands of dollars. For most people, RadeonClaw is a developer or enterprise story, not a home-lab story.
AMD’s “Agent PC” Vision — and the Mac Mini Comparison
AMD is framing this around a new category it calls the Agent Computer — a dedicated secondary machine running AI agents 24/7 while your main PC handles normal tasks. It’s the same conceptual territory as Perplexity’s Personal Computer announcement and the original OpenClaw on Mac mini approach that kickstarted this whole trend.
The honest comparison: Mac Minis have anecdotally become a popular choice for OpenClaw, combining Apple M-series silicon with a compact, power-efficient form factor. AMD’s pitch is that at 128GB, the Ryzen AI Max+ can handle significantly larger models than a base Mac mini — but the price premium is real and the setup is more involved.
For more on the local AI hardware race and what it means for everyday users, follow our AI and tech coverage on Technosports. And for the best AMD hardware to buy in India right now, check our AMD hardware guide on Technosports.
