AMD GPU Monitor

If you run ComfyUI on a Radeon card, you already know the annoying part: most of the "here's your GPU load and VRAM" tooling in this ecosystem was written for Nvidia and just doesn't see your hardware. Crystools - the utility most people reach for by default, right up there with ComfyUI Manager and rgthree on the "must-install" lists - pulls its numbers from nvidia-smi/pynvml, so on an AMD box its usage panel typically comes up blank or missing. AMD GPU Monitor exists purely to patch that hole: it's a small, single-purpose widget that reads your AMD GPU the way it actually reports data, and puts utilization, VRAM, and temperature in the corner of your screen where you can see them while a workflow runs.

How it works

No cleverness here, and that's the point. The extension shells out to rocm-smi (AMD's own command-line monitoring tool, sometimes amd-smi on newer stacks), parses what it prints, and renders it as a small floating panel - draggable, collapsible, closable, and it remembers where you left it between sessions. Color coding runs blue → orange → red as usage climbs, so you can tell at a glance whether you're comfortably under budget or one node away from an OOM. It doesn't touch inference, doesn't hook into the model loader, doesn't call any API - it's just polling a CLI tool on a timer and drawing the result. The author says outright it was inspired by Crystools' monitoring panel and built with Claude's help; it's not trying to be a framework, just the one panel AMD users were missing.

The inputs and outputs that matter

There's genuinely one dial here:

• update_interval - a float, default 1, adjustable from 0.1 to 10 seconds. This is how often the underlying node polls rocm-smi for a fresh reading. Lower it if you want tighter-grained numbers during a benchmark run; raise it if you'd rather not spawn a subprocess ten times a second on a modest machine.

The node has one STRING output - a text snapshot of the current reading. It's not marked as an output node, so on its own it does nothing visible; the floating panel you actually look at is injected automatically by the pack's JS side and doesn't require you to touch the graph at all. Per the README, "no setup is required" - install it, restart, and the panel just shows up. The node with update_interval and the string output is there if you want to wire a GPU-stats snapshot into something else, like a Show Text or Save Text node, so you get a logged reading alongside whatever you generated. Most people will never drag this node onto their canvas and will still get the full monitor.

Installing it

Standard custom-node install, nothing exotic:

cd ComfyUI/custom_nodes
git clone https://github.com/iDAPPA/ComfyUI-AMDGPUMonitor.git

Restart ComfyUI and the panel appears top-right. No Python packages to install, no model weights to download - it's a thin wrapper around a system tool you already need for AMD compute anyway. The real requirement isn't pip, it's your OS: you need rocm-smi (or amd-smi) installed and on PATH, which means ROCm drivers actually working, which means Linux. The README is specific about this - tested on Radeon RX 7900 XTX, Linux with ROCm. If you're on Windows, this one isn't for you: even though ComfyUI Desktop picked up official AMD ROCm support in January 2026, that path doesn't ship rocm-smi the way a Linux ROCm install does, and this pack has no Windows fallback documented.

Where people get burned

Almost every failure here traces back to ROCm itself, not the node. Getting rocm-smi and the ROCm stack working on Linux is a known source of pain independent of ComfyUI - driver package mismatches, missing render/video group membership for your user, kernel/driver version skew - and the community threads on setting it up are long for a reason. So the README's own troubleshooting order is the right one: run rocm-smi in a plain terminal first. If it doesn't return sane numbers there, the panel inside ComfyUI was never going to work either, and the fix is on the driver side, not this extension. If rocm-smi works standalone but the panel still stays blank, check the browser console for JS errors before anything else - it's a small, actively-vibecoded utility from one author rather than a maintained framework, so if AMD changes rocm-smi's output format in some future driver release, don't be shocked if parsing breaks until someone patches it. Worth keeping in mind: this replaces Crystools' monitoring panel specifically for AMD cards, it isn't a general Crystools substitute, so if you were relying on Crystools for other things (node execution timers, workflow stats), you'll still want that pack installed alongside this one.