June 27, 2026
Two PCs, one AI ego
AMD Strix Halo RDMA Cluster Setup Guide
DIY AI supercomputer guide drops, and the comments are already planning bigger, weirder builds
TLDR: A new guide shows hobbyists how to connect two AMD desktop boards into a much bigger AI setup using a super-fast link. Commenters are torn between awe at the DIY ambition and side-eye at the cost, with some already planning even larger homebrew clusters.
A new setup guide showing how to link two AMD Strix Halo desktop boards into one mini AI powerhouse has sent the community into full mad-scientist mode. The basic idea is surprisingly simple: connect two pricey little machines with a super-fast cable so they can act like one bigger brain for running giant language models. The guide is detailed, practical, and very much for the brave. But in the comments, the vibe is less “nice documentation” and more “I’m building a lab monster in my garage immediately.”
The loudest reaction is pure hype. One commenter called the memory-and-speed combo “amazing for homelabbers,” while another said this kind of work helps ordinary enthusiasts inch closer to the kind of AI setups usually reserved for giant companies. There’s also a heavy undercurrent of open-source hero worship, with shoutouts to container tools and side projects that make these Frankenstein builds possible.
But of course, the party got a reality check. The biggest hot take? This isn’t exactly cheap. One commenter did the back-of-the-napkin math and basically said: cool guide, but you’re still spending thousands, and the fancy network card can’t even fully stretch its legs in the available slot. That sparked the classic enthusiast split: is this a breakthrough for regular people, or just an expensive science fair for rich nerds?
The funniest energy comes from people already dreaming bigger: three-node clusters, Apple envy, and the unspoken meme hovering over the thread—normal people buy a desktop, these commenters are apparently building a basement data center.
Key Points
- •The article provides a setup guide for a two-node AMD Strix Halo cluster connected with Intel E810 RoCE v2 networking to run distributed vLLM inference.
- •The quick-start instructs users to configure Fedora 43, BIOS and kernel settings, passwordless SSH, static IPs, MTU 9000, and firewall trust on both nodes.
- •The `refresh_toolbox.sh` script is described as automatically creating an RDMA-capable container and applying a custom `librccl.so` patch.
- •The architecture uses vLLM for inference, Ray for cluster orchestration, RCCL for tensor synchronization, and RoCE v2 RDMA for low-latency memory transfer between nodes.
- •The guide specifies hardware including two Framework Desktop Mainboards with AMD Ryzen AI MAX+ Strix Halo, Intel E810-CQDA1-class 100GbE NICs, and a DAC cable, while noting the need for a PCIe riser due to the motherboard’s x4 slot.