The best GPU for ComfyUI is the one with enough VRAM for the model you run. Speed comes second. A 12 GB RTX 3060 makes a FLUX.1-dev image in about 64 seconds and an RTX 5090 in under 9, but both finish the job. The tiers differ in what fits on the card: 8 GB covers SDXL and the new distilled models, 16 GB streams FLUX at a usable pace, and 24 GB is where video models get comfortable.
So you get one pick per VRAM tier below, each with the SDXL and FLUX time to expect. We timed four of those cards in ComfyUI v0.42.1 on rented hosts, and published benchmarks cover the cards we didn’t rent, with the source on every number. Prices and stock for every pick were re-checked on Amazon US in November 2026, after three of the five October picks sold out.
Pick a GPU for ComfyUI by VRAM tier
Start with the tier that holds your model, then compare the price with the next tier up. Times are warm seconds per 1024×1024 image: SDXL at 30 steps and FLUX.1-dev fp8 at 20 steps, from our own runs.
| VRAM | Pick | Runs well | SDXL / FLUX | Price | Buy |
|---|---|---|---|---|---|
| 8 GB | ASUS Prime RTX 5060 8GB | SDXL, FLUX.3 Klein, Wan 2.2 5B with offload | not published for this card | about $535 | Check price |
| 12 GB | GIGABYTE RTX 3060 12GB | SDXL, FLUX fp8 streamed | 20.75 s / 63.91 s | about $460 | Check price |
| 16 GB | ASUS Dual RTX 5060 Ti 16GB | SDXL, FLUX fp8 streamed, Wan 14B at Q4 (tight) | 11.34 s / 37.02 s | about $850 | Check price |
| 24 GB | MSI RTX 3090 Ventus 3X OC | FLUX fp8 whole, Wan 14B at Q4 with headroom | 8.28 s / 23.23 s | about $2,500 | Check price |
| 32 GB | ASUS TUF RTX 5090 32GB | everything here, fastest | 4.84 s / 8.84 s | about $7,500 | Check price |
For most of you, the 16 GB RTX 5060 Ti is the card to buy. It runs every image workflow here in the default mode, it’s 1.7x to 1.8x faster than the 3060, and it costs about a third of the 3090 listing. Drop to the 3060 only when your budget stops under $500. Move to 24 GB when video is the reason you’re buying. We measured the 5090 as the fastest card by far, but at almost four times its $1,999 launch price it only pays off when time on the card is billable.
Compare SDXL and FLUX speed per card
Every host ran the same ComfyUI version, Docker image and torch build, so our rows compare directly. Each figure is the warm mean of three images after the model loaded, timed by the ComfyUI execution timer. The it/s value is the last sampler line in the server log:
| 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 |
FLUX widens the gaps you see on SDXL. The 5090 is 1.71x the 3090 on SDXL but 2.63x on FLUX. The ComfyUI GPU buying notes list 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 read it as the likely reason. Our host also held the 5090 at 450 W, below the 575 W NVIDIA rates it for, so a stock card should do a little better.
We did not rent an RTX 4090 or an 8 GB card for these runs. The 4090, 5080 and 4060 rows come from logs other users posted, and the 3060 Ti row is ours from a separate FLUX.3 Klein test. They used different ComfyUI builds and settings, so keep them apart from the table above:
| GPU | Result | Test | Source |
|---|---|---|---|
| RTX 4090 24GB | 11.28 s, 1.85 it/s | FLUX.1-dev fp8 template, 20 steps, sage attention | ComfyUI discussion 9002 |
| RTX 3090 24GB | 26 s | same thread and template | ComfyUI discussion 9002 |
| RTX 5080 16GB | 16.70 s, 1.33 it/s | same thread, stock template | ComfyUI discussion 9002 |
| RTX 4090 24GB | 7.46 s per image | Procyon SDXL FP16, 100 steps, Windows | StorageReview |
| RTX 5090 32GB | 5.22 s per image | same Procyon test | StorageReview |
| RTX 4060 8GB | 15.9 s, 1.72 it/s | 20 steps, per the thread’s default SDXL workflow | ComfyUI discussion 2970 |
| RTX 3060 Ti 8GB | 19.6 s, 1.82 it/s | FLUX.3 Klein int8, 16 steps, new prompt | our FLUX.3 Klein test |
In that ComfyUI thread, the 4090 made a FLUX image about 2.3x faster than the 3090 (11.28 s against 26 s). On the StorageReview Procyon run it sits about 30% behind the 5090 on SDXL. Our 5090 FLUX time of 8.84 s also lands within a tenth of a second of the 8.78 s a user posted there, which is a fair check on both. Each community row is a single log, though, so you should read them as a range, not a lab result.
Match VRAM to SDXL, FLUX and video models
Current ComfyUI turns on DynamicVRAM by itself and streams the weights that do not fit. A small card rarely fails outright; it slows down. So the real question is how much streaming you will put up with, and that depends on the model:
- 8 GB: SDXL, FLUX.3 Klein int8, Wan 2.2 5B with offloading
- 12 GB: FLUX.1-dev fp8, streamed, about a minute per image
- 16 GB: FLUX fp8 1.7x faster than 12 GB, Wan 14B at Q4_K_M and 480p with little to spare
- 24 GB: FLUX fp8 with nothing streamed, Wan 14B at Q4_K_M with room left
- 32 GB: everything above, with room for LoRAs and an upscaler
Our FLUX.3 Klein test backs the 8 GB line. So do the Wan 2.2 notes from ComfyUI, which say the 5B model “should fit well on 8GB vram with the ComfyUI native offloading”. The video rows lean on our Wan video guide, measured on a 24 GB RTX 4090. There, the quantized 14B model peaked at 14,907 MiB at 480p, so a 16 GB card should hold it with little to spare; we have not run it on one. The 5B model at 720p hit 24,079 MiB through the official Wan script. Even with its offload flags on, it ran out of memory at the VAE decode until the expandable_segments allocator was set. Video is where VRAM runs out first, so if you plan on it, 24 GB is the safer start and 32 GB the comfortable one.
We also forced the memory flags on the 12 GB and 16 GB cards to find where they break. --lowvram changed nothing, and the ComfyUI flag help explains why: the flag does nothing while dynamic VRAM is on. --novram ran FLUX in about 3 GB of VRAM, 15% slower on the 3060 and 22% slower on the 5060 Ti. --gpu-only keeps every weight on the card, and it ran out of memory on both. Here is the error from 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 MiB
A 16 GB card cannot hold the whole fp8 FLUX model, so it streams it, and the default mode does that without being told. Guides that tell you to add --lowvram after an out-of-memory error are out of date on current ComfyUI. That includes our own ComfyUI install guide, which still covers the install well, but you can start ComfyUI without that flag.
Give FLUX enough system RAM
System RAM mattered more than we expected. Our first RTX 3060 host had only 15 GB, 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). The first image, though, took 159.8 s instead of 68.3 s, and every new prompt cost 13 s more. That small host also had a slower PCIe Gen2 slot, so RAM is the main suspect but not the only one. With --gpu-only or --novram on the small host, 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 anything moves to the card. A 16 GB machine has little left for that once the OS, a browser and the ComfyUI server take their share. The kernel then swaps 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. Budget for it when you price the card, because a cheap GPU in a 16 GB box feels far slower than our table says.
Choose the card for your budget
Every card below costs well above its launch price, and our GPU price tracker tracks that gap. Amazon sells three of the five picks itself. The 3090 and the 5090 come from marketplace sellers, and we flag that where it matters.
ASUS Prime RTX 5060 8GB: the 8 GB card for SDXL
You buy into the 8 GB tier for SDXL and the new distilled models. The RTX 5060 is the current-generation card in it that Amazon sells itself.
Expect to pay about $535 for the ASUS Prime RTX 5060 8GB OC (check the live price on Amazon), well over its $299 launch price. We found no published ComfyUI timings for this exact card. The nearest evidence is two older 8 GB cards: an RTX 4060 posted 15.9 s for SDXL at 20 steps, and the RTX 3060 Ti in our FLUX.3 Klein test made a new-prompt image in 19.6 s. The 5060 is a Blackwell card with fp8 support, so we expect it near those figures, but that is an estimate.
Buy it if the PC also games, or if your case only takes a 145 W card on a 550 W supply. It’s 268 mm long and 2.5 slots thick. Skip it if FLUX.1-dev or video is the goal. The 12 GB RTX 3060 below costs about $75 less and holds 4 GB more, and in ComfyUI that VRAM matters more than the newer chip.
GIGABYTE RTX 3060 12GB: the cheapest card that runs FLUX
We ran FLUX.1-dev fp8 on nothing smaller than the 3060, and it’s still the cheapest way you can get 12 GB.
Amazon still sells the GIGABYTE RTX 3060 WINDFORCE OC 12G for about $460 (see the current price), 1.4x its launch price. In our runs it made a FLUX image in 63.91 s and an SDXL image in 20.75 s, about 56 FLUX images an hour. That’s slow but usable when one person works through prompts. The 170 W card is 198 mm long, so it fits almost any case.
The next tier is the catch. For about $390 more, the 5060 Ti runs FLUX 1.73x faster with 4 GB more VRAM. If you will run FLUX every day, the 3060 is the card you’ll replace first. Buy it when your budget stops under $500, and pair it with at least 32 GB of system RAM so cold loads don’t eat the savings.
ASUS Dual RTX 5060 Ti 16GB: best value for ComfyUI
Most readers should buy this one. You get 16 GB, a measured 1.7x over the 3060, and a 180 W card that a mid-range PC can power.
We linked the black ASUS Dual RTX 5060 Ti 16GB in October, and it has sold out. This pick is the white version of the same Dual OC card at about $850 (check the price now). Our 5060 Ti made a FLUX image in 37.02 s and an SDXL image in 11.34 s, 1.73x and 1.83x the 3060. It drew about 180 W under load and used less energy per FLUX image than the 3060 or the 3090.
You pay about 34% of the 3090 listing price for about 63% of its FLUX speed, the best ratio on this page. ASUS asks for a 550 W supply and NVIDIA for 600 W, so plan on 600 W. Skip it if you need FLUX held whole on the card or plan serious video work, because 16 GB streams the first and leaves little headroom for the second. If you also run local LLMs, our 5060 Ti vs 5070 Ti comparison shows how it does there.
MSI RTX 3090 Ventus 3X OC: 24 GB for video
You buy 24 GB for what it holds, not for speed: the whole FLUX fp8 model, quantized Wan 14B with room left, and bf16 models such as FLUX.2 Klein 9B, which Runpod measured at a 21.1 GB peak.
The MSI RTX 3090 Ventus 3X 24G OC is listed as New at about $2,500 from a marketplace seller and ships in 7 to 8 days (see the listing). Most other 3090 listings we found were Renewed titles or Used offers. The New exceptions were an EVGA FTW3 Ultra near $2,300, whose specs EVGA no longer publishes, and a Founders Edition near $2,900. The listing title names both GIGABYTE and MSI, so confirm the model line says Ventus 3X 24G OC. Our 3090 made a FLUX image in 23.23 s at its stock 350 W and kept the whole model on the card, peaking at 19,460 MiB. MSI rates this card at 350 W with two 8-pin plugs and a 750 W supply, and it is 305 mm long.
Used cards cost less than this listing, and the next section covers them. Take this listing if you want a 3090 sold as new through Amazon, and pay about $900 over the used average for it knowingly. Planning a second card for LLMs later? Our dual RTX 3090 build shows the power and slot math.
ASUS TUF RTX 5090 32GB: fastest, at a reseller price
We measured the 5090 as the fastest card by a wide margin. Its 32 GB also leaves you room for a video model plus LoRAs and an upscaler in one workflow.
Every new 5090 in our Amazon US search came from marketplace sellers. The ASUS TUF Gaming RTX 5090 OC runs about $7,500 from one of them, in low stock (check the live listing), 3.75x its launch price. Our capped 5090 made a FLUX image in 8.84 s, about 407 an hour. At about 1.06 Wh per image it also used the least energy of any card we ran.
Buy it if generation time is billable and you need 32 GB. The TUF card is 348 mm long and 3.6 slots thick, and ASUS asks for a 1,000 W supply. Check your case and PSU before the card; our PSU guide for AI builds covers the 12V-2×6 cable. Everyone else can get a used 4090 for under half the price, give up about 30% on SDXL, and keep 24 GB.
Buy a used RTX 3090 or 4090 for Stable Diffusion
For 24 GB on a budget, you’ll find the value in the used market, but in 2026 it isn’t cheap. The ResalePrices tracker follows US listings, and in early November it had both cards above their launch prices:
| Card | Used average | Listings seen | Change in 90 days | Launch price |
|---|---|---|---|---|
| RTX 3090 24GB | about $1,573 ($1,300 to $1,900) | 328 | up 25.9% | $1,499 |
| RTX 4090 24GB | about $3,471 ($2,850 to $4,500) | 27 | up 38.8% | $1,599 |
Those are listing-based averages, not checked sales, but the trend is plain. A used 3090 now costs about what a new one did at launch. A used 4090 costs more than twice its launch price, because RTX 40 cards are out of production and memory shortages pushed new cards up too. New 4090s on Amazon start around $4,800 from marketplace sellers, so buying used is the only sane way to get one.
Pick the 3090 when 24 GB is the point and you can live with its speed. It holds FLUX whole, we measured it ourselves, and at about $1,573 used it costs less than half a used 4090. Pick the 4090 when you’d otherwise look at a 5090. In the ComfyUI thread above it ran FLUX about 2x faster than our 3090, with sage attention on, and it has the fp8 support the 3090 lacks. NVIDIA lists 850 W of system power on its 4090 spec page and 750 W for the 3090 Founders Edition on its 3090 page.
A used card has no reviews to lean on, so test it before the return window closes. Run these checks in order:
- Keep proof of the sale from the seller
- Run memtest_vulkan for at least 6 minutes
- Run a long FLUX batch and watch power and temperature
- Seat the 4090 16-pin plug fully, with 35 mm before any bend
Warranty is the part buyers miss. The NVIDIA warranty page says partner cards are covered by the partner. The EVGA policy covers only the first buyer unless the card has its separate transferable warranty, which needs proof of the sale between you and the previous owner and excludes mining damage. memtest_vulkan ends with “testing PASSed” when the VRAM is clean. The 35 mm bend rule comes from the CableMod 12VHPWR guidance, written after the 4090 connector failures. If a seller won’t let you run these tests, skip the seller.
Skip these cards for now
Some cards look like obvious picks but don’t make sense at current prices, or Amazon itself didn’t sell them when we checked:
| Card | Why we left it out |
|---|---|
| RTX 5070 Ti 16GB | The PNY Slim we linked sold out, and the marketplace listings we checked were unavailable too |
| RTX 5080 16GB | About $2,300 for the same 16 GB as a 5060 Ti |
| RTX 4090 24GB, new | Marketplace sellers only, about $4,800 to $5,200 |
| RTX 5060 Ti 8GB | About $548, more than the RTX 5060 8GB for the same VRAM |
| Radeon AI PRO R9700 32GB | The ASRock card sold out; other R9700s are about $2,970 to $3,180 from resellers |
| Intel Arc Pro B60 24GB | About $1,000 and in stock, but we found no ComfyUI numbers for it |
Watch the 5070 Ti. Back2Gaming measured it at 9.84 SDXL images a minute in ComfyUI, about 6.1 s each, with the R9700 just behind at 9.03. Either one becomes a fair pick once Amazon stocks it again near the old price. For the 32 GB cards on the LLM side, our Arc Pro B70 vs R9700 comparison has the numbers.
Buy for the model you run next
Image models keep getting smaller and video models keep getting bigger, so size the card for where your work is heading:
- Stills only: 16 GB, the RTX 5060 Ti
- Video soon: 24 GB, a used 3090 you can test
- Video for paying work: 32 GB, the RTX 5090
Video models eat VRAM fastest, so a 16 GB card bought for stills is the first thing you outgrow if video follows. If the same card will also serve language models, our GPU picks for LLMs and the Ollama models by VRAM cheat sheet show what each tier runs there.




