https://nostr.ae https://nostr.ae/npub1vfacxn2wepmqguu93takp8x5w0a2drjluqvrkx8p4u47dv46mzasvxznvn https://nostr.ae/nevent1qqsp9km3m4dz7tmn4pjjg3f8ccudnu95clz3xcp6ln6emwedhw2g6eczyp38hq6dfmy8vprnsk90kcyu63el4f5wtlspswccuxhjhe4jhtvtkkqt0sd In reply to <a href='/nevent1qqszq57jyzyy34sa25cuhj8xq06m56nwmxep53syllwqgqagqlgu9jgwwa94u'>nevent1q…a94u</a><br/>_________________________<br/><br/>I&#39;ll note on that Ryzen AI Halo? It&#39;s Memory bandwidth is 256Gbs, so it&#39;s operating at a level on par with Framework&#39;s Max 395+ Desktop.<br/>That being said, for local AI compute, your options for Frontier Models is to either network multiple devices like this together through stuff like exolab&#39;s software, or to get something like a Threadripper/EPYC, or Intel equiv. to get together a bunch of old Server GPUs like V100s from Nvidia or the AMD options with a couple of adapter cards if you can&#39;t find a central server unit that the SXM interface expects.<br/>At that point, we&#39;re less getting a model together for AI Waifu Coom sessions, and building out a complicated stack for Homelab tier of AI compute, which means we&#39;re either exploring some niche use cases for AI, or building a bespoke coding assistant, or such. 2026-07-22T05:57:04+02:00