Does Image Generation Run on an Intel Arc B580? Trying It Without Touching the Host

2026年8月20日

15 August 2026

Everything I have put on the Intel Arc B580 so far has been text generation — LLMs. Chasing down why it was slow, measuring what fits inside 12GB, running it alongside an RTX 3090.

This time it is image generation. What I could find said “it apparently runs, but you get stuck" — and all of it was a while ago, with nothing recent. Would it run now? That is what I set out to find.

This is a record of running it here. The work was done on 15 August 2026.

The previous article is here.

What I wanted to do, and why I hesitated

The goal is simple: get image generation running on the B580. If it runs, compare it against the RTX 3090 sitting in the same desktop.

The comparison is for next time; this article is about getting there. Finding the right combination was where the real climb was.

To use the B580 for compute you need a compute runtime, separate from the driver that puts pixels on the screen. The one that ships with Ubuntu 24.04 is too old for the B580 generation.

There is a way to install a newer one. Intel and Canonical publish a repository jointly, and installing from it solves the problem. Except that the repository itself says it lacks the usual quality assurance and should not go on production machines.

This desktop is a working machine — it generates the images for this blog every day. I cannot afford to break it. That is where I sat for a while.

How I tried it: without touching the PC, inside a container only

Reading further, it turned out that what you need comes in two storeys.

So I shut the first floor inside a container. A container is an isolated box built inside the desktop. Whatever goes in the box, the outside is unaffected.

If it went wrong, deleting the box puts everything back. And in fact, across all of this work, nothing was added to the desktop itself.

The detour I found while reading up

Working through the instructions, one thing stood out.

Running machine learning on Intel GPUs used to mean adding an extension called IPEX. That is no longer needed. Development ended in March 2026, and PyTorch itself now supports Intel GPUs directly.

This is where a search can trip you up. Most of the instructions you find date from when IPEX was required, and following them sends you looking for a part that no longer exists.

I also found an installer someone had published, but that too had been retired — “ComfyUI supports this officially now, so this has served its purpose." Nobody needs to build the detour any more.

The combination that finally worked

This is the shape it ran in.

What the box can reach is deliberately narrow. Only the Intel GPU is passed through; the NVIDIA side is not. The model directory is mounted read-only — losing 249GB of models to an accident is not recoverable.

I checked recognition in four steps, in order: is the GPU visible, can PyTorch use it, does a matrix multiply run, do the operations image generation needs go through. Working down the list tells you immediately where it stopped.

It went through.

Straight after startup, the images came out pure black

With it running, I set it generating. That is where I hit a fault.

The first two images after startup came out pure black. The files were only 4KB, and inspecting them showed brightness at exactly zero. From the third onwards they were fine.

I retried several times, changing the order and the random seed, and it has not reproduced since. It seems to be something that happens in the first few images after startup.

The awkward part is that no error appears. As far as the screen is concerned it succeeded, and the files exist. Generate dozens in a batch without checking the contents and you would never notice.

If you use a B580 for this, it is worth checking the contents of the first few images as a matter of habit.

Confirming the B580 really was doing the drawing

Something nagged at me here.

What if the 3090 had been doing it? Both cards are in the same desktop, and if I had them mixed up, every comparison that follows is worthless.

I checked it two ways.

First, it is not reachable in the first place. Neither the NVIDIA GPU nor its driver is passed into the box. PyTorch itself is a build without NVIDIA support, and it duly reports “NVIDIA unavailable / Intel available". It could not use it if it tried.

Second, I watched the power draw of both cards while it was generating.

Only the Intel card raises its draw. The NVIDIA card never moves, and its utilisation stays at 0% throughout.

The CPU-did-it explanation is out too. CPU utilisation peaked at 31%, and what takes minutes on CPU alone was finishing in seconds.

How far this goes: what this article cannot tell you

What was established here is only whether it runs.

• Speed was not measured. The side-by-side against the RTX 3090 is the next article
• Image quality was not examined either. Whether the same settings produce the same picture is the article after that
• Only one model, SDXL, was tried. Whether other models behave the same way is unknown
• Only one card was used. Other Arc cards, and other distributions, are untested

In summary: image generation does run on a B580

• Image generation ran on the B580. Nothing was installed on the desktop itself; it is all contained inside the box
• The only thing missing was the compute runtime. The kernel side was already in place
• The IPEX extension is no longer needed. It ended in March 2026 and PyTorch took over. Copying older instructions sends you on a detour
• The first two images after startup came out black. No error is raised, so checking the first few is the safer habit
• Watching the power draw of both cards confirmed the B580 was the one computing

While reading up, everything said “it is supposed to work but you get stuck", and honestly I was not expecting much. Doing it, the sticking point turned out to be a single thing: the runtime being one version behind. Past that, it simply ran.

So it runs — but how fast? The next article puts it beside an RTX 3090 under matched conditions.

Hardware used in this test

Test environment: Ubuntu 24.04.4 / kernel 7.0.0-28 / Intel Arc B580 12GB / ComfyUI v0.16.4. Carried out 15 August 2026.