Three months wrong about why my 4-node AMD cluster was slow

Summary

The main topic is the powerful performance of a cluster of AMD Ryzen AI Max Plus 395 (Strix Halo) chips for AI tasks. Key subjects include the 128GB unified memory, X86 APU capabilities, and running the DeepSeek-R1 model. The practical takeaway is that a DIY cluster of these chips can outperform vendor demonstrations for AI inference, achieving impressive token generation speeds.