The EVO-X1: one box, three compute engines, no cloud
Problem
Cloud inference meant paying per token for work that touches source code and private data, plus latency that breaks the feel of a live coding session. We wanted a workstation that could host, build and run its own AI tooling — without shipping anything off-box.
Approach
We measured the real constraint first: CPU, iGPU and NPU all share one 120 GB/s memory bus, so decode speed is a bandwidth budget, not a raw compute one. We pinned clocks, benchmarked model and quantization combinations against that budget, and wired a model router with memory guardrails so only one model ever holds the GPU at a time.
Outcome
The box now hosts and builds itself: a coding agent, a knowledge base, and the render pipeline for this site, all running local, revertible, and reachable only over a private Tailscale mesh — no standing cloud dependency for the work that matters most.
Links
Related
Every local-AI build starts with the same question: how fast can it actually go? On a unified-memory machine the honest answer is a bandwidth budget, not a spec sheet — CPU, integrated GPU and NPU all draw from the same memory bus, so the real law governing decode speed is bandwidth divided by bytes per token.
We treated that law as the design brief. Clock pinning came first — an unpinned fabric clock quietly halves decode throughput and is the first thing worth checking whenever a run feels slow. From there we benchmarked model and quantization pairings against the measured ceiling instead of a marketing number, and built a small router that keeps exactly one model resident at a time so memory never gets over-committed.
What came out the other side is a workstation that hosts its own tooling: an AI coding agent, a searchable knowledge base, and the very pipeline that builds and serves this site — all local-first, reachable only over a private network, and revertible at every layer. Nothing about the work that touches this studio's source or client data has to leave the building.