A 12 GB RTX 3060 runs FLUX.1-dev in ComfyUI. It takes about 64 seconds per image, and an RTX 5090 takes under 9. We tested ComfyUI v0.42.1 on four cards across five rented hosts on October 20, 2026, with the same SDXL and FLUX workflows, fixed seeds and ComfyUI’s own timer on every run. That gives you real per-card numbers to set against today’s inflated GPU prices.
We also pushed the 12 GB and 16 GB cards with ComfyUI’s memory flags to see where they break. The answer changes some common advice about the best GPU for ComfyUI, because the default mode now streams weights on its own.
| GPU | SDXL, 30 steps | FLUX.1-dev fp8, 20 steps | FLUX images per hour | FLUX peak VRAM |
|---|---|---|---|---|
| RTX 3060 12GB | 20.75 s (1.53 it/s) | 63.91 s (0.33 it/s) | about 56 | 11,857 MiB |
| RTX 5060 Ti 16GB | 11.34 s (2.88 it/s) | 37.02 s (0.55 it/s) | about 97 | 15,749 MiB |
| RTX 3090 24GB | 8.28 s (4.16 it/s) | 23.23 s (0.89 it/s) | about 155 | 19,460 MiB |
| RTX 5090 32GB, capped at 450 W | 4.84 s (7.10 it/s) | 8.84 s (2.38 it/s) | about 407 | 17,638 MiB |
Check the cards we measured
Every number you see here used 1024×1024, batch 1 and the warm mean of three images after the models were loaded. SDXL was the base model with euler at 30 steps. FLUX was ComfyUI’s fp8 dev checkpoint at 20 steps. Each card ran on its own vast.ai host with the same Docker image and torch build.
On FLUX the gap is wider than on SDXL. The 5090 is 1.71x the 3090 on SDXL but 2.63x on FLUX. ComfyUI lists fp8 support for the 40 and 50 series only, which fits the 3090 losing more ground on an fp8 checkpoint. We did not isolate that in a test, so treat it as the likely reason.
Read the 5090 row as a floor for what you’d get from a stock card. That host capped the card at 450 W of its 600 W default. For the same two cards on LLM work, our RTX 3090 vs 5090 comparison has the tokens per second.
Pick the VRAM for the model you run
FLUX.1-dev in fp8 ran to completion on both 12 GB and 16 GB in ComfyUI’s default mode. Current ComfyUI turns on DynamicVRAM by itself and streams the weights that don’t fit. The 24 GB and 32 GB cards had room for the whole model, and FLUX peaked at 17,638 MiB on the 5090.
We then tried the three memory flags on the smaller cards. This is what each one did:
- –gpu-only: out of memory on 12 GB and 16 GB
- –lowvram: no change at all
- –novram: FLUX in about 3 GB of VRAM
ComfyUI’s own flag help explains the –lowvram result: it says the flag does not do anything when dynamic VRAM is enabled. So older guides that tell 12 GB owners to add it are out of date. With –novram the 3060 took 73.52 s per FLUX image, 15% slower, and the 5060 Ti took 45.15 s, 22% slower.
You’ll meet the –gpu-only failure if you force everything onto a 16 GB card. This was the exact error on the 5060 Ti:
torch.OutOfMemoryError: Allocation on device 0 would exceed allowed memory. (out of memory)
Currently allocated : 15.23 GiB
Requested : 40.00 MiB
Device limit : 15.48 GiB
Free (according to CUDA): 11.06 MiBIf you want video, that is a bigger job than still images, and we did not rerun it here. Our Wan video guide has measured VRAM figures for Wan 2.2 on a 24 GB card, and the FLUX guide covers the model itself.
Get enough system RAM for FLUX
System RAM turned out to matter more than we expected. Our first RTX 3060 host had only 15 GB of RAM, so we reran the same card on a host with about 150 GB. Warm FLUX speed barely moved, 65.15 s against 63.91 s.
Everything else got slower on the small host. The first FLUX image took 159.8 s instead of 68.3 s, and a new prompt cost 13 s more. With –gpu-only or –novram, ComfyUI failed before it touched the GPU:
RuntimeError: unable to mmap 17246524772 bytes from file </workspace/cfg/ComfyUI/models/checkpoints/flux1-dev-fp8.safetensors>: Cannot allocate memory (12)On disk the FLUX fp8 checkpoint is 17.2 GB, and ComfyUI maps it into system memory before it moves anything to the card. A 16 GB machine has little room left for it once the OS, the browser and the ComfyUI server take their share, so the kernel starts swapping or refuses the mapping. From those two hosts, we’d put 32 GB of system RAM as the floor for FLUX-class models and 64 GB as comfortable. That small host also had a slower PCIe Gen2 slot, so the two causes are not fully separated.
Choose the card for your budget
Prices below are from Amazon US on October 20, 2026, and every card here costs well above its launch price. Our GPU price tracker follows the gap week by week. The first three cards are ones we measured. The two faster 16 GB cards use third-party results, and we say whose. If you want the short version, start here:
- Under $500: RTX 3060 12GB
- Best value: RTX 5060 Ti 16GB
- 24 GB for video: renewed RTX 3090
- Fastest 16 GB: RTX 5070 Ti or RTX 5080
ASUS Dual RTX 5060 Ti 16GB: best value
At $765.99, the ASUS Dual RTX 5060 Ti 16GB OC is the card we’d recommend first. In our lab it made a FLUX image in 37.02 s and an SDXL image in 11.34 s, 1.73x and 1.83x the speed of the 3060. It drew about 180 W under load and used less energy per FLUX image than the 3060 or the 3090. It costs about 40% of a renewed 3090 and gives you about 63% of its FLUX speed. It also runs both workflows in default mode with no flags, so you lose only time, never the ability to run the model.
ASUS asks for a 550 W power supply and NVIDIA says 600 W, so plan on 600 W. The card uses one 8-pin connector and is 229 mm long. If you also run local LLMs, our 5070 Ti vs 5060 Ti test shows how it does there.
GIGABYTE RTX 3060 12GB: the cheapest card that runs FLUX
At $459.99, the GIGABYTE RTX 3060 WINDFORCE OC 12G costs 1.4x its launch price. It made a FLUX image in 63.91 s and an SDXL image in 20.75 s, about 56 FLUX images an hour. That is slow but usable, and it is the cheapest way we found into 12 GB.
Price is the catch. The 5060 Ti costs $306 more and is about 1.7x faster on FLUX, with 4 GB more VRAM. Buy the 3060 if your budget stops under $500, and stay with SDXL or FLUX fp8. Our Stable Diffusion install guide sets up ComfyUI on Linux for either card.
NVIDIA RTX 3090 Founders Edition (Renewed): 24 GB that holds FLUX whole
As a Renewed unit, the RTX 3090 Founders Edition is $1,899.99, and only a few were left. It made a FLUX image in 23.23 s at its stock 350 W, and it held the whole model in VRAM. That 24 GB is the reason to buy it, not raw speed.
A renewed 3090 now costs about the same as a new RTX 5080, so this is the real decision at this budget. Third-party SDXL tests put the 5080 ahead of the 5070 Ti, which in turn beat our 3090 result. But with 16 GB it has to stream FLUX the way the 5060 Ti did in our lab, while the 3090 keeps the whole model on the card. Pick the 3090 if you run video models or bf16 checkpoints that need more than 16 GB, and the 5080 if you only make still images. Check the Renewed return terms on the listing, and see our dual RTX 3090 build if you plan a second card later.
PNY RTX 5070 Ti Slim OC: a faster 16 GB card
PNY sells its RTX 5070 Ti Slim OC for $1,299.99. The only 5070 Ti host within our rental budget refused our SSH key, so these figures are third-party. Back2Gaming measured 9.84 SDXL images a minute in ComfyUI at 1024×1024 and 30 steps, about 6.1 s each.
On SDXL that puts it between our 3090 and 5090 results, with 16 GB that still streams FLUX. StorageReview’s 100-step Procyon SDXL test on Windows gave it 11.18 s per image, against 20.41 s for the 5060 Ti. Those runs use more steps and a different engine, so compare them with each other and not with our table. The Slim model is 290 mm long and two slots wide, and it needs a 750 W supply and one 16-pin cable.
ASUS ProArt RTX 5080 OC: the fastest 16 GB card
For $1,869.46 you get the ASUS ProArt RTX 5080 OC, which we did not test ourselves. StorageReview measured 8.81 s per Procyon SDXL image, which is 1.27x faster than the 5070 Ti on the same test. It has the same 16 GB, so FLUX still streams part of its weights.
Buy it if you want the most speed you can get without going past 16 GB. The ProArt is 304 mm long and 2.5 slots wide, and ASUS asks for an 850 W supply. If VRAM matters more than speed, the renewed 3090 costs the same.
Pass on these cards for now
A few cards look like obvious picks but don’t make sense at today’s prices, or couldn’t be bought when we checked:
| Card | Why we left it out |
|---|---|
| RTX 5090 32GB | Fastest we measured, but every new 5090 on Amazon came from a third-party seller at about $7,900 or more |
| RTX 4090 24GB | New units from third-party sellers only, at about $4,450 to $4,700 |
| RTX 5070 12GB | $855.99, more than the 16 GB 5060 Ti |
| Radeon AI PRO R9700 32GB | Not tested by us; the ASRock card at $1,699.99 was listed as shipping later on Amazon |
Intel’s Arc Pro B70 is the cheaper 32 GB card at $1,299.99, though Amazon also listed it as shipping later. Back2Gaming measured the R9700 at 9.03 SDXL images a minute in ComfyUI on Linux, a little behind the 5070 Ti. Puget Systems also tested it on Z-Image Turbo. Our Arc Pro B70 vs R9700 comparison covers its LLM side.
Start ComfyUI with the right flag
Start ComfyUI with no memory flag first. That’s the mode that ran FLUX on every card we tested. On any NVIDIA 20 series or newer card, the ComfyUI README also says you need a cu130 or newer PyTorch build. From the ComfyUI folder, run:
python main.pyOnly add –novram if you need the VRAM for something else, such as a second model or a game. It cut FLUX to about 3 GB on both small cards, for a 15 to 22% slowdown, but it still needs the system RAM covered above:
python main.py --novramIn the server log you can see the weights moving off the card. This was the line on the 5060 Ti:
[INFO] loaded partially; 0.00 MB usable, 0.00 MB loaded, 11350.07 MB offloaded, 315.10 MB buffer reserved, lowvram patches: 0Leave –lowvram out on NVIDIA cards, since it does nothing there in the current ComfyUI release, and do not use –gpu-only on anything under 24 GB. If you’re still choosing between image work and LLMs, our VRAM guide for LLMs and GPU picks for LLMs cover the other side.
