{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreif2yjgebgdwt3zjc4p747qladpdisdi6tmxtdivmqdvefw4raqayi",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mjnrkfmeegz2"
  },
  "path": "/t/gpu-advice-on-5080/175204#post_12",
  "publishedAt": "2026-04-17T00:02:36.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "amd.com",
    "pytorch.org",
    "rocm.docs.amd.com",
    "nvidia.com"
  ],
  "textContent": "I think it’s best to be extra cautious when considering buying a GPU, especially with prices skyrocketing right now…\n\nTo put it simply, “It’s an **XFX** card with an **AMD GPU** , and since it’s a **old-generation model** , it’s cheap”:\n\n* * *\n\nNo. It is **not cheap because it is a bad card**.\n\nIt is cheap because it sits in an awkward but interesting spot:\n\n  * **very strong hardware**\n  * **older generation**\n  * **24GB VRAM**\n  * **AMD, not NVIDIA**\n  * **Windows AI setup is still less smooth than CUDA**\n\n\n\nThat combination pushes the price down.\n\n## The short version\n\n**Why cheap?**\nBecause the market discounts **AMD on Windows** , not because the hardware is weak.\n\n**Bad card?**\nNo. The hardware is still strong.\n\n**Good for your use?**\nPotentially yes, but only if you accept more setup friction than with NVIDIA.\n\n* * *\n\n## What the card actually is\n\nThe **RX 7900 XTX** is still a serious GPU.\n\nAMD’s official specs list:\n\n  * **24GB GDDR6**\n  * **384-bit bus**\n  * up to **960 GB/s** memory bandwidth\n  * recommended **800W PSU** (amd.com)\n\n\n\nThat is a lot of VRAM.\nFor local AI image work, that part is genuinely attractive.\n\nSo this is **not** a weak card pretending to be high-end.\n\n* * *\n\n## Why it is cheaper than the NVIDIA cards\n\n### 1. It is older\n\nThe 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 So in 2026 it is no longer current-generation halo hardware.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)\n\n### 2. NVIDIA gets the AI convenience premium\n\nFor 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)\n\n### 3. AMD on Windows is better than before, but still less polished\n\nAMD’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)\n\n### 4. XFX is usually a value-focused board partner\n\nXFX 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)\n\n* * *\n\n## So is it a bad card?\n\n**No.**\n\nIt is better described as:\n\n**good hardware + weaker software ecosystem on Windows + lower market demand for AI buyers**\n\nThat is why it is cheap.\n\nIf this same card had **CUDA** and the same software support level as NVIDIA, it would almost certainly cost more.\n\n* * *\n\n## What the real trade-off is\n\nYou are not choosing between:\n\n  * **good card**\n  * **bad card**\n\n\n\nYou are choosing between:\n\n  * **better hardware value**\n  * **easier Windows experience**\n\n\n\nThat is the real decision.\n\n### If you buy the XFX 7900 XTX\n\nYou get:\n\n  * **24GB VRAM**\n  * strong raw hardware\n  * very good value for money\n\n\n\nBut you also accept:\n\n  * AMD/Windows quirks\n  * a less beginner-friendly path\n  * more chance of setup friction with ComfyUI and Hugging Face\n\n\n\n### If you buy NVIDIA instead\n\nYou usually get:\n\n  * less VRAM for the money\n  * worse raw value\n  * but a smoother and more beginner-friendly Windows path\n\n\n\n* * *\n\n## My tier view\n\n### Hardware tier\n\n**A tier**\n\nThe hardware itself is strong.\n\n### Value tier\n\n**A / S tier**\n\nIf the price is much lower than comparable NVIDIA options, the value can be excellent.\n\n### Beginner-Windows tier\n\n**B tier**\n\nNot because the card is bad.\nBecause the **Windows software path** is less relaxed than NVIDIA.\n\n### Reputation tier\n\n**B+ / A-**\n\nXFX is not a brand I would call low-tier or suspicious.\nIt is more “good value” than “luxury premium.”\n\n* * *\n\n## My honest answer for you\n\nIf you want:\n\n### the easiest possible Windows experience\n\nThis is **not** the best choice.\n\n### the most VRAM for the money\n\nThis is **one of the most interesting choices**.\n\n### the safest beginner route\n\nI would still lean NVIDIA.\n\n### the best raw hardware bargain\n\nThe XFX 7900 XTX is hard to ignore.\n\n* * *\n\n## Final verdict\n\n**Why cheap?**\nOlder card, AMD discount, no CUDA premium, value-oriented board partner, and weaker Windows AI convenience.\n\n**Bad card?**\nNo.\n\n**Good buy?**\nYes, **if** you care more about **24GB + value** than about the easiest Windows setup.\n\n**Best beginner choice?**\nNo.\nIt is a **value-first choice** , not the **lowest-friction choice**.\n\n* * *\n\nThe 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)\n\nQuestion / Attribute | RTX 5070 Ti 16GB | RTX 5080 16GB | RX 7900 XTX 24GB\n---|---|---|---\nVRAM / memory path | 16GB GDDR7, 256-bit. (nvidia.com) | 16GB GDDR7, 10,752 CUDA cores. (nvidia.com) | 24GB GDDR6, 384-bit, up to 960 GB/s. (amd.com)\nIf you want the **least hassle** on Windows | Good | **Best** | Weakest of the three\nIf you want the **best value** | **Best** | Good, but pricier | Good if you specifically value 24GB\nIf you want the **most headroom** | Fair | Good | **Best**\nIf you are a **beginner** using ComfyUI + HF on Windows | **Best** | Very good | Riskiest\nIf you hate paying the “ecosystem premium” | Best | Fair | **Best for hardware value**\nMy blunt answer | Buy if budget matters | Buy if you want the cleanest strong upgrade | Buy only if you knowingly accept more setup friction\n\n## The same decision tree in plain English\n\nYour answer | Pick\n---|---\n“I want the cheapest card that still feels like a big upgrade.” | **5070 Ti**\n“I want the easiest strong Windows upgrade.” | **5080**\n“I want the most VRAM for the money, even if setup is a bit more annoying.” | **7900 XTX**\n“I am new and I do not want to debug the platform.” | **5070 Ti or 5080**\n“I care more about long-term headroom than convenience.” | **7900 XTX**\n“I want the least regret overall.” | **5080**\n\n## Online reputation tier vs price tier\n\nThis 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.”\n\nVendor / line style | Online reputation tier | Price tier | What that usually means\n---|---|---|---\n**Sapphire Pulse / Nitro+** | **A** | **A to B** | Usually trusted AMD picks. Often sensible rather than absurd.\n**PowerColor Hellhound** | **A** | **A** | Good Radeon reputation without going full halo.\n**PowerColor Red Devil** | **A** | **C** | Strong reputation, but often halo-priced.\n**XFX MERC / Speedster** | **B+ / A-** | **S / A value** | Usually the value-first AMD choice. Often cheaper than prestige AMD brands.\n**MSI Ventus / Gaming Trio** | **A-** | **A / B** | Usually balanced Nvidia picks. Easy to recommend when price is sane.\n**Gigabyte Windforce / Gaming** | **B+ / A-** | **A / S value** | Often good value. Reputation is decent, not usually the “prestige tax” brand.\n**Zotac Solid / Trinity** | **B+** | **S / A value** | Often aggressively priced. Good when you want the GPU tier more than the premium brand.\n**Palit / Gainward** | **B** | **S value** | Usually value brands. Good on price, less “aspirational” reputation.\n**ASUS TUF / ProArt** | **A** | **B / C** | Good reputation, but often priced above the sweet spot.\n**ASUS ROG Strix / Astral** | **A** | **C halo** | Premium and usually expensive. Often more luxury than value.\n\n## The shortest summary\n\n**Best overall:** **5080**\n**Best value:** **5070 Ti**\n**Best VRAM value:** **7900 XTX**\n\nAnd for board partners:\n\n**Best “safe without silly pricing” brands:** Sapphire, Hellhound-class PowerColor, MSI mid/high lines.\n**Best value brands:** XFX, Zotac, Gigabyte, Palit, Gainward.\n**Most likely to be overpriced:** halo ASUS, Red Devil-class halo cards, other flagship-tier boards.",
  "title": "GPU advice on 5080"
}