# GPU advice on 5080 **URL:** \ **Category:** Beginners\ **Created:** [April 13, 2026, 7:42am UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204 "2026-04-13T07:42:59Z")\ **Posts on this page:** 14\ **Page:** 1
**Author:** ![kdalts3759](https://avatars.discourse-cdn.com/v4/letter/k/b782af/32.png) [@kdalts3759](https://discuss.huggingface.co/u/kdalts3759)\ **Post date:** [April 13, 2026, 7:42am UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/1 "2026-04-13T07:42:59Z")
Hi, I hope this post is in the correct place… just started my journey using Hugging Face and Comfy so please be patient. It is evident that my current GPU (4060 8GB VRAM) is not man enough for the task… even the most simple workflows like edit an image can take 5 to 10 mins. For some reason the price and availability for a high end card in the UK has gone crazy… Amazon with cards at £5k? The best I can narrow it down to is a Gigabyte GeForce RTX 5080 WINDFORCE OC SFF 16G Graphics Card - 16GB GDDR7, 256bit, PCI-E 5.0, 2670MHz Core Clock, 3 x DP 2.1a, 1 x HDMI 2.1b, GV-N5080WF3OC-16GD Could someone advise it this would be a worthwhile upgrade as copilot is saying that the 50xx series is not a good choice for this kind of task Many thanks in advance ---
**Author:** ![bacca400](https://avatars.discourse-cdn.com/v4/letter/b/ee59a6/32.png) [@bacca400](https://discuss.huggingface.co/u/bacca400)\ **Post date:** [April 13, 2026, 9:28am UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/2 "2026-04-13T09:28:56Z")
IMO wait a few months for RAM and VRAM prices to go down. Prices went up because the news claimed there were contracts to buy 12 months worth of 90% of production of VRAM and SSDs from major manufacturers. Well, OpenAI did not have a contract, it had a letter of intent, which it has recently broken. In West Virginia monthly power bills are now more than peoples’ mortgages due to a local data center. Cities attract data centers to claim they created jobs but the regular people pay the bill. I think we’ll see more resistance to data centers, letters of intent will not become contracts for a few companies, and VRAM and SSD prices will come down. Here are some sites with price histories per unit of RAM/VRAM. 1. [RAM Price History | Historical DRAM Prices 1957-Present](https://mfara.com/unlimited/) 2. PC Part Picker. [https://pcpartpicker.com/trends/price/memory/](https://pcpartpicker.com/trends/price/memory/) 3. RAM Scout. [https://pcpartpicker.com/trends/price/memory/](https://pcpartpicker.com/trends/price/memory/) ---
**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\ **Post date:** [April 13, 2026, 10:50am UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/3 "2026-04-13T10:50:23Z")
True. GPU prices are skyrocketing all over the world, though the extent varies by country. The surge in prices for VRAM and RAM, plus HDDs, SSDs is really painful…🥶 It’s gotten to the point where they’re actually remanufacturing older-generation GPUs (like the 3060 Ti)… Even Moore would be surprised. That aside, the reason Copilot treats the RTX 5080 like a landmine is that, until about halfway through last year, it actually was one. If Copilot draws on its built-in knowledge or picks up old articles via web searches, it will likely conclude that “Blackwell is still too early for AI applications.” In reality, there’s absolutely no problem if you’re using ComfyUI **today**. The real concern is VRAM capacity. While it’s absolutely true that “16GB is better than 8GB,” that doesn’t necessarily mean “16GB is enough.” If you prioritize running large, modern models smoothly over peak performance with smaller models, there’s a clear incentive to choose a card with more VRAM—even if it’s from a slower, older generation. Also, while it barely runs on Windows 10 and is still advanced-user territory on Windows 11, in a Linux environment, AMD GPUs become a realistic option for ComfyUI. If you search for terms like “Wan 2.2 AMD ComfyUI,” you’ll find plenty of how-to articles. However, for the time being, I can state with certainty that “if you’re using open-source AI models, an **NVIDIA GPU is definitely the easier choice** \*.” * * * For your use, **VRAM** , **Windows software maturity** , and **price-to-performance** matter more than factory clock speed. ComfyUI’s own docs show that once you are pushed into `lowvram` or `novram` behavior, speed and usability drop fast, and your RTX 4060 is exactly the kind of card that gets pushed there: **8GB GDDR6, 128-bit, 3,072 CUDA cores**. By contrast, the RTX 5080 class gives you **16GB GDDR7, 256-bit, and far more compute** , which is why this upgrade feels so different in practice. ([ComfyUI](https://docs.comfy.org/interface/settings/server-config "Server Config")) ## Best overall The RTX 5080 gets called “bad” mostly for two reasons. First, early Blackwell support was rough, so a lot of launch-era advice got stuck in people’s heads. PyTorch only added official Blackwell support in **2.7** with **CUDA 12.8** wheels in April 2025. Current ComfyUI guidance is now much more normal: the Windows portable Nvidia build uses **Python 3.13 + CUDA 13.0** , which tells you the platform is no longer being treated as a fragile special case. So on a **clean 2026 install** , the old “50-series is not good for this” claim is largely outdated. ([PyTorch](https://pytorch.org/blog/pytorch-2-7/ "PyTorch 2.7 Release")) For **your purpose** , the 5080 is good. Your exact Gigabyte card is **16GB GDDR7, 256-bit, 10,752 CUDA cores, 2670 MHz core clock** , and Gigabyte specifies **304 × 126 × 50 mm** , **850W recommended PSU** , and a **single 16-pin power connector**. That is a major jump from your 4060, and it is exactly the kind of jump that makes image generation and image editing stop feeling like a constant memory-management exercise. ([GIGABYTE](https://www.gigabyte.com/Graphics-Card/GV-N5080WF3OC-16GD "GeForce RTX™ 5080 WINDFORCE OC SFF 16G")) It is not perfect. The 5080 is still a **16GB** card, so it is not the “I never think about VRAM again” option. Benchmark-wise, it is clearly faster than a 5070 Ti in AI image generation, but it still trails a 4090 in relevant Stable Diffusion tests. In StorageReview’s Procyon results, the 5080 scored **4,650** in Stable Diffusion 1.5 FP16, while the 4090 scored **5,260**. That is why the 5080 is best described as **strong and clean** , not **ultimate**. ([StorageReview.com](https://www.storagereview.com/review/nvidia-geforce-rtx-5080-review-the-sweet-spot-for-ai-workloads "NVIDIA GeForce RTX 5080 Review: The Sweet Spot for AI ...")) ## Best value This is the option I would call the smartest spend. The 5070 Ti still gives you the part that matters most: **16GB GDDR7 on a 256-bit bus**. Gigabyte’s own spec page shows **8,960 CUDA cores** and a **2497 MHz** factory clock for this exact Windforce card. In AI image benchmarks, it is slower than the 5080, but not weak. StorageReview measured Stable Diffusion 1.5 FP16 at **1.664 seconds per image** on the 5070 Ti versus **1.344 seconds** on the 5080. So the 5080 is faster, but the 5070 Ti already gets you out of the painful 8GB zone. Current product results in this chat show the 5070 Ti class is materially cheaper than the 5080 class, which is why I call it the value winner. ([GIGABYTE](https://www.gigabyte.com/Graphics-Card/GV-N507TWF3OC-16GD "GeForce RTX™ 5070 Ti WINDFORCE OC SFF 16G")) ## Best alternative if you care more about headroom than Windows simplicity The appeal here is obvious: **24GB VRAM** and a **384-bit bus**. That gives you much more breathing room for heavier models and future experimentation. AMD’s official RX 7900 XTX page confirms the card line, and board specs from major vendors show the expected **24GB GDDR6** and **384-bit** memory path. The catch is software maturity on Windows. AMD’s ROCm Windows compatibility matrix does include the RX 7900 XTX, but ComfyUI’s own system requirements still describe AMD Windows/Linux support as **experimental** for the relevant Radeon generations. So for a beginner on Windows, I would treat this as a **specialist alternative** , not the default safe choice. On Linux, it becomes much more attractive. ([AMD](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html "Radeon™ RX 7900 XTX")) | Attribute | Gigabyte GeForce RTX 5080 WINDFORCE OC SFF 16GB | **Gigabyte GeForce RTX 5070 Ti WINDFORCE OC SFF 16GB** | **Sapphire Radeon RX 7900 XTX Pulse 24GB** | | --- | --- | --- | --- | | VRAM / bus | 16GB GDDR7, 256-bit ([GIGABYTE](https://www.gigabyte.com/jp/Graphics-Card/GV-N5080WF3OC-16GD/sp "GeForce RTX™ 5080 WINDFORCE OC SFF 16G")) | 16GB GDDR7, 256-bit ([GIGABYTE](https://www.gigabyte.com/Graphics-Card/GV-N507TWF3OC-16GD "GeForce RTX™ 5070 Ti WINDFORCE OC SFF 16G")) | 24GB GDDR6, 384-bit ([AMD](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html "Radeon™ RX 7900 XTX")) | | Current Windows support picture | Mainstream Nvidia path in current PyTorch + ComfyUI ([PyTorch](https://pytorch.org/blog/pytorch-2-7/ "PyTorch 2.7 Release")) | Mainstream Nvidia path in current PyTorch + ComfyUI ([PyTorch](https://pytorch.org/blog/pytorch-2-7/ "PyTorch 2.7 Release")) | Supported in ROCm Windows matrix, but AMD path is still experimental in ComfyUI ([ROCm Documentation](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/windows/windows_compatibility.html "Windows support matrices by ROCm version")) | | AI-image speed signal | Faster than 5070 Ti, but still behind 4090 in cited SD tests ([StorageReview.com](https://www.storagereview.com/review/nvidia-geforce-rtx-5080-review-the-sweet-spot-for-ai-workloads "NVIDIA GeForce RTX 5080 Review: The Sweet Spot for AI ...")) | Respectable, but slower than 5080 in SD 1.5 FP16 ([StorageReview.com](https://www.storagereview.com/review/asus-prime-geforce-rtx-5070-ti-review "ASUS Prime GeForce RTX 5070 Ti Review")) | Main attraction is memory headroom rather than the smoothest Windows benchmark story in the cited sources ([AMD](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html "Radeon™ RX 7900 XTX")) | | Best fit | Best overall for a beginner Windows setup | Best value | Best if 24GB matters more than convenience | | Main downside | Same 16GB ceiling as cheaper cards | Slower than 5080 | More Windows-side friction than Nvidia | My answer for **your exact case** is this: the **5080 is not a bad choice in 2026**. It is actually a **good** one. The old bad reputation mostly came from early Blackwell support friction and from people seeing inflated prices and expecting 24GB-class flexibility from a 16GB card. If you want the **best overall** card for Windows + ComfyUI + Hugging Face with the least hassle, the 5080 is the right answer. If you want the **best value** , the 5070 Ti is the sharper buy. If you know you will care more about future VRAM headroom than Windows smoothness, then 24GB cards become interesting, but they are not the safest first move for a newcomer. ([PyTorch](https://pytorch.org/blog/pytorch-2-7/ "PyTorch 2.7 Release")) One last rule of thumb, because it really matters here: for your kind of work, the order is **VRAM first** , then **software support/maturity** , then **memory bus and GPU tier** , and **clock speed last**. The difference between 8GB and 16GB changes your daily experience far more than a small clock uplift ever will. ([ComfyUI](https://docs.comfy.org/interface/settings/server-config "Server Config")) ---
**Author:** ![kdalts3759](https://avatars.discourse-cdn.com/v4/letter/k/b782af/32.png) [@kdalts3759](https://discuss.huggingface.co/u/kdalts3759)\ **Post date:** [April 13, 2026, 1:42pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/5 "2026-04-13T13:42:26Z")
very comprehensive reply, many thanks indeed… given I will probably not want to be buying another card after this in the near future, it sounds like a 24GB is the way to go.. not AMD… would that be a fair assessment? Then it’s about cost… maybe, as suggested, wait for prices to drop ---
**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\ **Post date:** [April 13, 2026, 8:24pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/6 "2026-04-13T20:24:17Z")
Yeah. Good choice. If you go with 24GB of VRAM, you’ll have significantly more headroom than with 16GB. As you know, the baseline is around 8GB, so from a headroom perspective, it’s effectively double… which is especially advantageous when doing complex tasks like ControlNet. It’s also beneficial for high-resolution generation. That said, well, the price is a bit of an issue… As for the price, it ultimately comes down to the market speculation mentioned above and the mismatch between semiconductor manufacturers’ production plans and actual demand. (Even if semiconductor manufacturers set production plans, they can’t suddenly change what they’re producing. There’s a long time lag between manufacturing and shipping. That has always led to wild price swings… (There were even times when DRAM could be bought for next to nothing!) So, I think prices will eventually come down, but personally, I’m expecting it to take several months or more… unless there’s some global-scale disruption (like new market entrants or a recession). I’m not good at market forecasting, so I’ll avoid making specific predictions… ---
**Author:** ![Pimpcat-AU](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/pimpcat-au/32/48989_2.png) [@Pimpcat-AU](https://discuss.huggingface.co/u/Pimpcat-AU)\ **Post date:** [April 14, 2026, 11:44pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/7 "2026-04-14T23:44:31Z")
I’ve been in tech for the majority of my adult life and I am old now. Never, EVER buy Gigabyte hardware… It is junk. When it comes to GPU’s stick to ASUS 1st or MSI as a second choice. Never waste cash on any other companies cards or chances are you’ll never get your moneys worth out of it. ASUS TUF is generally the best option. Just make sure your case has optimized airflow. ---
**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\ **Post date:** [April 15, 2026, 12:42am UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/8 "2026-04-15T00:42:57Z")
By the way, since my PC case is a bit tight (in terms of GPU slot length), I’m an MSI fan when it comes to GeForce cards. ---
**Author:** ![kdalts3759](https://avatars.discourse-cdn.com/v4/letter/k/b782af/32.png) [@kdalts3759](https://discuss.huggingface.co/u/kdalts3759)\ **Post date:** [April 16, 2026, 12:50pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/9 "2026-04-16T12:50:48Z")
thanks for the advice… it’s a big investment so trying to make the right choice isn’t easy… some reports say to stay away from AMD… do you agree? ---
**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\ **Post date:** [April 16, 2026, 1:34pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/10 "2026-04-16T13:34:50Z")
First off, it’s not that AMD cards have performance issues. **The problem lies in driver and library support, which varies significantly from one OS to another**. While things are improving day by day, almost no one really knows where we stand right now… That’s why so many people say, “I wouldn’t actively recommend it.” On Windows 10, it’s quite a gamble. On Windows 11, if you’re just using ComfyUI, you’ll probably be fine. I’m sure of it. On Linux, ROCm support has matured to the point where it’s almost on par with CUDA… or so I’ve heard. I’m a Windows user, so this is secondhand information. As for Mac, I have no idea… Many people consider AMD as an option for the same reason you do. This is especially true for those using ComfyUI with open-source video generation models (like Wan, Hunyuan, or LTX). That’s because the difference between 16GB and 24GB is absolutely critical. So, there should be plenty of how-to articles from last year and this year. If video generation models work, image generation models usually run even more smoothly (excluding minor plugins). Therefore, I think the best approach is to start by looking for those kinds of articles… ---
**Author:** ![kdalts3759](https://avatars.discourse-cdn.com/v4/letter/k/b782af/32.png) [@kdalts3759](https://discuss.huggingface.co/u/kdalts3759)\ **Post date:** [April 16, 2026, 9:42pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/11 "2026-04-16T21:42:47Z")
sorry to keep coming back to this but there is so much conflicting information out there… This looks too good to be true.. XFX Speedster MERC310 AMD Radeon RX 7900XTX Black Gaming Graphics Card with 24GB GDDR6, AMD RDNA 3 RX-79XMERCB9… half the price of the intel versions… which are not available in the UK as far as I can see… what do you think? ---
**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\ **Post date:** [April 17, 2026, 12:02am UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/12 "2026-04-17T00:02:36Z")
I think it’s best to be extra cautious when considering buying a GPU, especially with prices skyrocketing right now… To put it simply, “It’s an **XFX** card with an **AMD GPU** , and since it’s a **old-generation model** , it’s cheap”: * * * No. It is **not cheap because it is a bad card**. It is cheap because it sits in an awkward but interesting spot: - **very strong hardware** - **older generation** - **24GB VRAM** - **AMD, not NVIDIA** - **Windows AI setup is still less smooth than CUDA** That combination pushes the price down. ## The short version **Why cheap?** Because the market discounts **AMD on Windows** , not because the hardware is weak. **Bad card?** No. The hardware is still strong. **Good for your use?** Potentially yes, but only if you accept more setup friction than with NVIDIA. * * * ## What the card actually is The **RX 7900 XTX** is still a serious GPU. AMD’s official specs list: - **24GB GDDR6** - **384-bit bus** - up to **960 GB/s** memory bandwidth - recommended **800W PSU** ([amd.com](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html)) That is a lot of VRAM. For local AI image work, that part is genuinely attractive. So this is **not** a weak card pretending to be high-end. * * * ## Why it is cheaper than the NVIDIA cards ### 1. It is older The RX 7900 XTX was announced in **November 2022** and started shipping in **December 2022**. So in 2026 it is no longer current-generation halo hardware. Older high-end cards often fall hard in price. ([amd.com](https://www.amd.com/en/newsroom/press-releases/2022-11-3-amd-unveils-world-s-most-advanced-gaming-graphics-.html)) ### 2. NVIDIA gets the AI convenience premium For Windows users doing local AI, NVIDIA is still easier because of **CUDA** and the broader “it just works” ecosystem around PyTorch and tools like ComfyUI. That convenience keeps NVIDIA prices high. ([pytorch.org](https://pytorch.org/blog/pytorch-2-7/)) ### 3. AMD on Windows is better than before, but still less polished AMD’s ROCm docs do include the **RX 7900 XTX** in the Windows-supported list. So it is not unsupported. But ComfyUI’s own system requirements still describe AMD support as **experimental**. That is the real catch. ([rocm.docs.amd.com](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/windows/windows_compatibility.html)) ### 4. XFX is usually a value-focused board partner XFX is an official AMD board partner, not a random or suspect brand. It tends to sit more on the **value/performance** side than on the prestige-premium side. AMD also notes that board partner availability varies by region, which fits your UK comment that some alternatives are hard to find. ([amd.com](https://www.amd.com/en/resources/support-articles/faqs/GPU-91.html)) * * * ## So is it a bad card? **No.** It is better described as: **good hardware + weaker software ecosystem on Windows + lower market demand for AI buyers** That is why it is cheap. If this same card had **CUDA** and the same software support level as NVIDIA, it would almost certainly cost more. * * * ## What the real trade-off is You are not choosing between: - **good card** - **bad card** You are choosing between: - **better hardware value** - **easier Windows experience** That is the real decision. ### If you buy the XFX 7900 XTX You get: - **24GB VRAM** - strong raw hardware - very good value for money But you also accept: - AMD/Windows quirks - a less beginner-friendly path - more chance of setup friction with ComfyUI and Hugging Face ### If you buy NVIDIA instead You usually get: - less VRAM for the money - worse raw value - but a smoother and more beginner-friendly Windows path * * * ## My tier view ### Hardware tier **A tier** The hardware itself is strong. ### Value tier **A / S tier** If the price is much lower than comparable NVIDIA options, the value can be excellent. ### Beginner-Windows tier **B tier** Not because the card is bad. Because the **Windows software path** is less relaxed than NVIDIA. ### Reputation tier **B+ / A-** XFX is not a brand I would call low-tier or suspicious. It is more “good value” than “luxury premium.” * * * ## My honest answer for you If you want: ### the easiest possible Windows experience This is **not** the best choice. ### the most VRAM for the money This is **one of the most interesting choices**. ### the safest beginner route I would still lean NVIDIA. ### the best raw hardware bargain The XFX 7900 XTX is hard to ignore. * * * ## Final verdict **Why cheap?** Older card, AMD discount, no CUDA premium, value-oriented board partner, and weaker Windows AI convenience. **Bad card?** No. **Good buy?** Yes, **if** you care more about **24GB + value** than about the easiest Windows setup. **Best beginner choice?** No. It is a **value-first choice** , not the **lowest-friction choice**. * * * The first table is the blunt **5070 Ti vs 5080 vs 7900 XTX** decision table for **your** use. The second table is the practical **online reputation tier vs price tier** view for board partners. The key facts behind the first table are simple: the 5070 Ti and 5080 are both **16GB** Nvidia cards on the current, well-supported Windows path, while the 7900 XTX gives you **24GB** and a **384-bit** bus but on a less turnkey Windows path because ComfyUI still describes AMD support as experimental. PyTorch 2.7 also matters because it is the point where Blackwell support became normal instead of awkward. ([pytorch.org](https://pytorch.org/blog/pytorch-2-7/)) | Question / Attribute | RTX 5070 Ti 16GB | RTX 5080 16GB | RX 7900 XTX 24GB | | --- | --- | --- | --- | | VRAM / memory path | 16GB GDDR7, 256-bit. ([nvidia.com](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/)) | 16GB GDDR7, 10,752 CUDA cores. ([nvidia.com](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/)) | 24GB GDDR6, 384-bit, up to 960 GB/s. ([amd.com](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html)) | | If you want the **least hassle** on Windows | Good | **Best** | Weakest of the three | | If you want the **best value** | **Best** | Good, but pricier | Good if you specifically value 24GB | | If you want the **most headroom** | Fair | Good | **Best** | | If you are a **beginner** using ComfyUI + HF on Windows | **Best** | Very good | Riskiest | | If you hate paying the “ecosystem premium” | Best | Fair | **Best for hardware value** | | My blunt answer | Buy if budget matters | Buy if you want the cleanest strong upgrade | Buy only if you knowingly accept more setup friction | ## The same decision tree in plain English | Your answer | Pick | | --- | --- | | “I want the cheapest card that still feels like a big upgrade.” | **5070 Ti** | | “I want the easiest strong Windows upgrade.” | **5080** | | “I want the most VRAM for the money, even if setup is a bit more annoying.” | **7900 XTX** | | “I am new and I do not want to debug the platform.” | **5070 Ti or 5080** | | “I care more about long-term headroom than convenience.” | **7900 XTX** | | “I want the least regret overall.” | **5080** | ## Online reputation tier vs price tier This second table is **practical opinion** , not failure-rate science. Public hard RMA/failure data by vendor is thin, inconsistent, or retailer-specific, so the useful version is “how enthusiasts usually treat them” plus “how they usually price themselves.” | Vendor / line style | Online reputation tier | Price tier | What that usually means | | --- | --- | --- | --- | | **Sapphire Pulse / Nitro+** | **A** | **A to B** | Usually trusted AMD picks. Often sensible rather than absurd. | | **PowerColor Hellhound** | **A** | **A** | Good Radeon reputation without going full halo. | | **PowerColor Red Devil** | **A** | **C** | Strong reputation, but often halo-priced. | | **XFX MERC / Speedster** | **B+ / A-** | **S / A value** | Usually the value-first AMD choice. Often cheaper than prestige AMD brands. | | **MSI Ventus / Gaming Trio** | **A-** | **A / B** | Usually balanced Nvidia picks. Easy to recommend when price is sane. | | **Gigabyte Windforce / Gaming** | **B+ / A-** | **A / S value** | Often good value. Reputation is decent, not usually the “prestige tax” brand. | | **Zotac Solid / Trinity** | **B+** | **S / A value** | Often aggressively priced. Good when you want the GPU tier more than the premium brand. | | **Palit / Gainward** | **B** | **S value** | Usually value brands. Good on price, less “aspirational” reputation. | | **ASUS TUF / ProArt** | **A** | **B / C** | Good reputation, but often priced above the sweet spot. | | **ASUS ROG Strix / Astral** | **A** | **C halo** | Premium and usually expensive. Often more luxury than value. | ## The shortest summary **Best overall:** **5080** **Best value:** **5070 Ti** **Best VRAM value:** **7900 XTX** And for board partners: **Best “safe without silly pricing” brands:** Sapphire, Hellhound-class PowerColor, MSI mid/high lines. **Best value brands:** XFX, Zotac, Gigabyte, Palit, Gainward. **Most likely to be overpriced:** halo ASUS, Red Devil-class halo cards, other flagship-tier boards. ---
**Author:** ![yotta-labs](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/yotta-labs/32/57526_2.png) [@yotta-labs](https://discuss.huggingface.co/u/yotta-labs)\ **Post date:** [April 20, 2026, 4:42pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/13 "2026-04-20T16:42:27Z")
Pricing has been pretty inconsistent lately depending on region and availability, not just for consumer GPUs but across cloud providers as well There are tools that aggregate pricing across providers and show ranges/trends, which can help give a better sense of where the market is at: > **[Live GPU Price Comparison: H100, A100, B200 Rental Costs | Mercatus](https://www.mercatus-ai.com/gpu-index)** > > Find the cheapest cloud GPU for your workload. Compare real-time hourly rental prices across 30+ platforms. Not exactly the same as buying a card, but useful context when thinking about overall GPU pricing ---
**Author:** ![kdalts3759](https://avatars.discourse-cdn.com/v4/letter/k/b782af/32.png) [@kdalts3759](https://discuss.huggingface.co/u/kdalts3759)\ **Post date:** [April 21, 2026, 5:44pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/14 "2026-04-21T17:44:52Z")
Thanks for all the replies… it’s a tricky issue given the insane prices right now… making the wrong choice could be an expensive mistake.. I found this on Amazon ## **NVIDIA RTX PRO 4000 Blackwell Graphics Card - 24GB GDDR7 ECC Memory for ai image to video generation** the spec sounds perfect but then other reports say the blackwell architecture is weak for AI image\>video generation I’m so confused - lol ---
**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\ **Post date:** [April 21, 2026, 10:42pm UTC](https://discuss.huggingface.co/t/gpu-advice-on-5080/175204/15 "2026-04-21T22:42:10Z")
Since it’s a really expensive purchase, I think it’s better to be cautious… It’s better than forgetting to ask after you’ve already bought it… > other reports say the blackwell architecture is weak for AI image\>video generation That report was accurate in the first half of last year, but **it’s now just a misunderstanding**. That’s because, at the time, ComfyUI and one of its core components (PyTorch) weren’t fully compatible yet. Blackwell is fundamentally a better architecture for AI than previous generations. While there are some poorly made Blackwell cards out there, that’s not unique to Blackwell… seriously. Also, the reason you’re confused is that “workstation and data center-grade cards”—which weren’t part of previous comparisons—have now been added as an option. All the GPUs we’ve considered so far have been designed for gaming. Gaming GPUs prioritize peak performance, consumer-friendly features, and flashy lighting effects. Workstation-grade cards, on the other hand, prioritize power efficiency, stability, computational accuracy, and durability. “If I could buy a card with the same price, performance, and VRAM—and if it fits in my PC—I’d choose a workstation-grade card without hesitation.” Since they’re more energy-efficient, they put less strain on your PC’s power supply than gaming cards. Normally, workstation-grade cards are extremely expensive, but… even if the processing speed drops slightly—say, by a few percentage points—they’re still a very attractive option. * * * What makes this confusing is that people often mix up **three different things** : the **architecture** , the **specific card** , and the **software stack**. Blackwell itself is not the real problem now: PyTorch 2.7 added Blackwell support, so the old “Blackwell is bad for AI” line is mostly launch-era baggage. The **RTX PRO 4000 Blackwell** looks attractive because it has **24GB GDDR7 ECC** , but it is also a **workstation card** built around lower power, compact form factor, and pro reliability rather than maximum consumer throughput. At the same time, image-to-video is a heavier workload than plain image generation because memory demand grows with **frames, resolution, and model size**. Wan 2.2’s official TI2V 5B model card explicitly frames a high-end consumer GPU as the natural reference point for 720p/24fps generation, which is why 24GB-class cards keep coming up in these conversations. ([pytorch.org](https://pytorch.org/blog/pytorch-2-7/)) ([nvidia.com](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-4000/)) ([huggingface.co](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B)) | Attribute | NVIDIA RTX PRO 4000 Blackwell | Gigabyte GeForce RTX 5080 GAMING OC SFF 16G | MSI GeForce RTX 5070 Ti 16G VENTUS 3X OC | Sapphire RX 7900 XTX Pulse Gaming OC 24GB | | --- | --- | --- | --- | --- | | Card type | Workstation / pro GPU. ([nvidia.com](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-4000/)) | Consumer GeForce. ([nvidia.com](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/)) | Consumer GeForce. ([nvidia.com](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/)) | Consumer Radeon. ([amd.com](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html)) | | Memory / special feature | 24GB GDDR7 **ECC** , workstation feature set, lower-power design. ([nvidia.com](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-4000/)) | 16GB GDDR7. ([nvidia.com](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/)) | 16GB GDDR7. ([nvidia.com](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/)) | 24GB GDDR6, 384-bit, up to 960 GB/s. ([amd.com](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html)) | | What it is really optimized for | Reliability, pro workflows, compact workstations, enterprise-style use. ([nvidia.com](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-4000/)) | Best balanced GeForce path for demanding local AI on Windows. | Best value GeForce step-up from 8GB. | Best raw VRAM-per-money, not best Windows simplicity. | | Main attraction for image-to-video | 24GB helps fit heavier jobs. | Strong Windows path, modern GeForce stack. ([pytorch.org](https://pytorch.org/blog/pytorch-2-7/)) | Much cheaper route into the 16GB class. | 24GB plus wide bus at a lower cost than equivalent 24GB GeForce options. | | Main catch | You are paying for workstation features, not just raw speed. | Still only 16GB, so heavier image-to-video jobs can hit the ceiling sooner. | Same 16GB ceiling, so easier to outgrow. | ComfyUI still describes AMD support as experimental on Windows. ([docs.comfy.org](https://docs.comfy.org/installation/system_requirements)) | | My blunt verdict | Good **only if** you want workstation benefits. | **Best overall**. | **Best value**. | **Best 24GB value** , but more fiddly on Windows. | The **RTX PRO 4000 Blackwell** is therefore **not a bad card**. It is just a **different kind of card**. It gives you workstation strengths: **ECC memory, pro drivers, lower power, compact size, and video engines**. Those things are useful in pro environments, multi-app reliability, or space-constrained systems. But for a normal desktop running **ComfyUI image-to-video** , they do not automatically make it the smartest buy. What matters more in that use case is usually **raw throughput per pound** plus **how smooth the Windows software path is**. ([nvidia.com](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-4000/)) ([pytorch.org](https://pytorch.org/blog/pytorch-2-7/)) That is why my answer changes depending on what you want: If you want the **simplest, strongest Windows choice** , pick the **RTX 5080**. It is the cleanest “buy it, install current drivers, use current ComfyUI and PyTorch, and get on with it” answer. The downside is simply that it is **16GB** , not 24GB. ([pytorch.org](https://pytorch.org/blog/pytorch-2-7/)) If you want the **cheapest sensible GeForce option** , pick the **RTX 5070 Ti**. It is the easiest card to justify when the current market feels insane, because it fixes the real pain of the 4060 8GB without pretending to be a forever card. ([nvidia.com](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/)) If you specifically want **24GB** but do **not** care about workstation extras, the **RTX PRO 4000 Blackwell** becomes harder to justify. It is very appealing on paper because “24GB GDDR7 ECC” sounds perfect, but the more honest reading is: **great reliability card, less obvious value card**. Unless you have a reason to want workstation traits, I would not make it the default recommendation for hobbyist or enthusiast ComfyUI image-to-video. ([nvidia.com](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-4000/)) If you want **24GB at the best raw value** , the **RX 7900 XTX** is the obvious alternative. The reason people keep bringing it up is simple: **24GB** and a **384-bit bus** are genuinely useful for heavier image and video workloads. The trade-off is that the Windows path is still less relaxed than GeForce. ([amd.com](https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html)) ([docs.comfy.org](https://docs.comfy.org/installation/system_requirements)) So the clean summary for **your purpose** is: - **Best overall:** RTX 5080 - **Best value:** RTX 5070 Ti - **Best workstation-only case:** RTX PRO 4000 Blackwell - **Best 24GB value alternative:** RX 7900 XTX And the plain-English answer to your original confusion is: **Blackwell is not weak here.** What is confusing people is the difference between a **workstation Blackwell card** and a **consumer Blackwell card**. The PRO 4000 is built to be **professional, stable, compact, and efficient**. The 5080 is built to be **faster for consumer-style GPU workloads**. For ComfyUI image-to-video, that distinction matters more than the architecture name. ([nvidia.com](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-4000/)) ([pytorch.org](https://pytorch.org/blog/pytorch-2-7/))