Image Generation on an Arc B580 vs an RTX 3090 — Same Picture, Measured for Speed and Power

2026年8月21日

15 August 2026

Last time I got image generation running on an Intel Arc B580. Once something runs, the next thing you want to know is how fast.

My desktop currently has a B580 and an RTX 3090 in it. One is 12GB at around ¥30,000 ($200); the other is 24GB and still goes for something like ¥200,000 ($1,340) second-hand. Give them the same picture at the same settings — how far apart do they land?

This is a record of measurements taken here, on 15 August 2026.

The previous article is here.

The test setup

The only thing I want to compare is the graphics card. Everything else is held the same.

Generation times are taken from the second image onwards, after the model is loaded. The first one has the load time mixed into it.

Results

Time per image

Generation itself was a draw. Not close — the same number, before any rounding. The only gap is in loading the model, where the 3090 finishes about eight seconds sooner.

The draw was not what I expected. While reading up I had seen several assessments saying the gap to NVIDIA was still wide. For SDXL under these conditions, at least, that did not hold.

The load-time gap matters little unless you swap models constantly. Start up and produce a few dozen images and it is eight seconds, once.

Power: the B580 draws about half of what the 3090 does

The gap turned up in power instead. 297.3W against 155.3W on average while generating. Per image that is 0.547Wh against 0.286Wh — roughly half.

In money, about ¥0.009 ($0.00006) per image. Ten thousand images comes to ¥89 (about $0.60), so this is not an electricity bill to brace for.

Where the power difference does bite is elsewhere. One is power-supply headroom: a card that pulls close to 300W and one that gets by on 150W allow different builds. The other is heat, which you notice running long jobs in summer.

What happens at higher resolutions

With speed and power level, the remaining difference had to be memory. 24GB against 12GB is a factor of two.

So I pushed the resolution and the batch size up, looking for where it falls over.

Not one of them fell over. The 12GB B580 got through 2048×2048 and a batch of eight.

Following the memory during generation showed why. When it runs short, the software moves parts of the model out to the desktop’s own RAM. It is not built to fail when memory runs out; it is built to shuttle data and keep going.

The extra time that shuttling costs was smaller than I expected. The widest gap was 1.3x, and on a batch of eight the B580 actually finished first.

“12GB cannot do high resolutions" did not hold, at least for SDXL. It gets there if you wait, and the wait grows by only 1.2–1.3x.

That said, all of this is with a model that fits inside 12GB (SDXL, 6.5GB). A larger model means more shuttling and a larger penalty. What happens with a big model needs checking separately.

Why no gap opened up in generation speed

Generating text, the B580 and the RTX 3090 were clearly apart. Here they are barely distinguishable. Same two cards — why does the answer change?

The reason is that different things set the pace for text and for images.

Put another way, which card suits you swaps around with what you are doing. The same two cards separate on text and tie on images. Judge from only one of those reputations and you will get it wrong.

What I could not put a number on during the runs

Memory usage could not be reported as a figure. I sampled every 0.5s, but the shuttling is violent enough that the peaks slip through. There were points where raising the resolution made the number appear to go down. All this article can say is whether it fell over, and how long it took.

Power, as noted above, was measured differently on the two cards. The Intel figure is exact; the NVIDIA one may read slightly low.

Only one model was tested. Everything here is SDXL at 6.5GB. Larger models are not covered.

In summary: a B580 is plenty to enjoy this on

• Generation speed tied with the RTX 3090 (SDXL 1024×1024, 20 steps — 6.0s on both)
• The only gap was loading the model, where the 3090 finishes about eight seconds sooner
• Energy per image is about half (0.286Wh against 0.547Wh). That shows up in power-supply headroom and heat, not the bill
• 12GB was not a wall. 2048×2048 and a batch of eight both went through
• What does not fit is moved out to the desktop’s RAM. The penalty was 1.2–1.3x the time
• Text separates them, images tie. Different things set the pace for each

So much for speed. But the important question is still open: are the pictures produced at identical settings actually identical? The next article puts them side by side.

Hardware used in this test

Test environment: Ubuntu 24.04.4 / kernel 7.0.0-28 / RTX 3090 24GB + Intel Arc B580 12GB (both in the same desktop) / ComfyUI v0.16.4. Measured 15 August 2026. Prices are Japanese retail; ¥1,000 is roughly $6.70.