gpuvrambuying-guidelocal-ai
How Much VRAM Do You Actually Need? UK GPU Guide
Confused about how much VRAM you need for gaming or local AI models? Here's a practical 2026 UK GPU buying guide, from budget 8GB cards to 16GB+ for LLMs.
Confused about how much VRAM you need for gaming or local AI models? Here's a practical 2026 UK GPU buying guide, from budget 8GB cards to 16GB+ for LLMs.
Every few months a thread pops up online where someone trains a model, runs a game, or edits a video on a GPU that's supposedly "too old" for the job — and it works fine. The reaction is always the same split: half the room insists 6GB or 8GB is obsolete, the other half points out they're doing real work on exactly that card right now. Both sides are talking past each other, because VRAM need depends entirely on what you're doing, and nobody ever says that part out loud.
So here's the straight version: how much VRAM you need for gaming is a different question from how much you need for local AI work, and conflating the two is why so much GPU-buying advice online is useless. Let's split them apart.
For gaming, VRAM requirements track resolution and texture settings far more than raw GPU horsepower does. As a working rule for 2026:
Undershoot these and the symptom isn't a crash — it's stuttering and texture pop-in as the card swaps assets in and out of memory faster than it can keep up, even though the GPU core itself has plenty of horsepower left. That mismatch is why a card can benchmark well in reviews and still feel rough in the one game you actually play with ultra textures.
Browse current cards by memory size on our GPU price tracker rather than relying on a spec sheet from six months ago — VRAM configurations on the same model number occasionally change between regions and revisions, and UK street prices move weekly.
Running or training generative models — image diffusion, small LLMs, audio and voice models — has completely different VRAM economics than gaming, and this is where most buying advice falls apart. A model that gaming benchmarks would call "VRAM-starved" can be perfectly viable for AI work, because you control batch size, resolution, and precision, and there's no fixed target frame rate you're failing to hit.
A recent example that did the rounds: someone trained a full generative audio diffusion model, from spectrograms through to usable samples, entirely on a seven-year-old 6GB gaming card in an old desktop. Not a cloud GPU, not a workstation card — a mid-range gaming GPU that's long since dropped off most people's radar. The discussion that followed was more useful than the project itself: a chunk of commenters insisted that hardware was hopelessly outdated for "real" AI work, while just as many pointed out plenty of people are doing genuinely useful generative work on 6-8GB cards right now, provided they size the model and batch to fit rather than assuming they need a 24GB card to get started.
Both things are true at once. VRAM headroom absolutely matters for AI work — it caps your batch size, your model size, and how much you can keep resident without swapping to system RAM (which tanks performance badly for training, less so for inference). But "6-8GB" and "useless" are not the same category. Here's a more honest breakdown:
If your interest is genuinely local AI rather than gaming, don't default to the biggest gaming flagship you can afford — check VRAM capacity first, since a mid-tier card with more memory will often serve you better than a higher-tier one with less. Our GPU listings let you sort by memory rather than just by name recognition, which is the number that actually matters here.
The honest answer, cutting through the online arguing: a 6GB card from several generations back is genuinely fine for smaller AI experiments and 1080p gaming, and genuinely limiting the moment you want to run anything resembling a modern mid-size model or game at 1440p+ with high textures. It's not a binary "obsolete" or "future-proof" — it's a real, specific ceiling that depends entirely on what you're pointing it at.
If you've already got a 6-8GB card sitting in a desktop, it's worth using it to figure out what you actually want to do — train a small model, run a quantised LLM locally, game at 1080p — before spending money upgrading. You'll learn a lot more about what VRAM tier you actually need from an afternoon of hands-on use than from any spec-sheet comparison. Only once you hit a genuine wall (out-of-memory errors, unbearable swap-induced slowdowns, textures visibly failing to load) does an upgrade become the obvious next move rather than a guess.
One mix-up worth clearing up: system RAM and VRAM are not interchangeable, even though some software will technically let a model spill over into system memory when VRAM runs out. That fallback works — technically — but the performance hit is severe, because you're moving data across the PCIe bus instead of keeping it on the GPU die. If you're planning serious local AI work, more system RAM is not a substitute for more VRAM; it's a separate, complementary upgrade. Check our RAM pricing guide if you're weighing where to put your budget, and our RAM category page for current UK prices — 32GB is a sensible baseline if you're doing any local AI work alongside gaming, since your OS, browser, and any dataset preprocessing all compete for that memory too.
If you're gaming only and staying at 1080p, don't overspend on VRAM you won't use — put the budget into the GPU core instead and check our GPU price tracker for the current best-value 8GB cards. If you're at 1440p or planning to be within the card's lifetime, treat 12GB as the minimum, not the ceiling.
If local AI is part of the plan — even casually, even just running a small model for fun — buy for VRAM capacity ahead of raw gaming benchmarks. A 12-16GB mid-range card will serve you better for that use case than a higher-clocked card with less memory, even if the latter wins gaming benchmarks by a few percent. And if you're not sure yet whether you'll actually use it, don't buy ahead of the need — a lot of people upgrade for AI work they end up not doing, when an afternoon on their existing 6-8GB card would have told them whether it was worth it at all.
Whatever tier you land on, cross-reference current listings on our full product catalogue before you buy — GPU pricing in the UK has been volatile enough this year that a card's relative value can shift within weeks, and last month's "obvious pick" isn't always this month's.
If you're buying specifically for local AI, one more thing shifts the numbers: quantisation. A model advertised as needing 16GB at full precision might run comfortably in 8-10GB once quantised to 4-bit (the GGUF and GPTQ formats you'll see referenced constantly in local-AI communities exist for exactly this reason). This is why you'll see wildly different VRAM requirements quoted for what's nominally "the same model" — one figure assumes full precision, the other assumes a compressed version. Quantisation trades a small amount of output quality for a large reduction in memory footprint, and for most hobbyist use cases that trade is easily worth it. Don't buy a bigger card than you need because you read one unquantised VRAM figure and took it as gospel.
How much VRAM do I need for 1440p gaming? Treat 12GB as the floor. Some newer titles at 1440p with ray tracing and high-res textures enabled will comfortably use more, so 16GB gives you genuine headroom rather than just scraping by.
How much VRAM do I need for local AI models? It depends entirely on model size and quantisation. Small quantised models run fine in 6-8GB; the 12-16GB range is where most hobbyists doing real local AI work land comfortably; 20GB+ is for larger models or serious fine-tuning.
Is an old 6GB card still worth using? For 1080p gaming and small-scale AI experiments, yes. It becomes a limitation the moment you want 1440p+ gaming or anything beyond small, quantised AI models — at which point the ceiling is real, not imagined.
Does more system RAM help if I'm low on VRAM? Only as an emergency fallback, and a slow one. It's not a substitute for buying enough VRAM in the first place — see our RAM category page if you need to top up system memory anyway, but budget for the GPU upgrade separately.
Source: How to Train a Gen AI Kick Drum Model on Your Old Linux Desktop with 6GB VRAM
Compare prices and find deals on the products mentioned in this article.
This article was written with the assistance of AI tools and reviewed by a human editor. Price data is sourced from Amazon UK. For more information, see our About page.