NVIDIA GeForce RTX 5080

The RTX 5080 is one of the fastest cards you can buy new, and for local LLMs it arrives with a frustrating asterisk: 16GB. The GDDR7 memory and Blackwell silicon are seriously quick, but the capacity is the same as cards costing a third as much. Great for models that fit, limiting for the ones that don’t.

The take

Buy it for blistering speed on models up to about 14B, on a warranty, if the 16GB ceiling suits your work. The 960 GB/s of GDDR7 is roughly the fastest inference bandwidth at this price, and it shows in every session. But at around £1000, paying flagship-adjacent money for 16GB stings when a used 3090 gives you 24GB for a third of that. This is a card you buy because it’s fast and new, not because it’s sensible VRAM value. If your models fit in 16GB, though, very few new cards feel quicker.

What it’ll actually run

Sixteen gigabytes comfortably runs 8B to 14B models at Q4 with a generous context, and it runs them very fast. A 32B at a tight quant will fit if you keep the context modest, but you’re managing memory rather than relaxing into it. The moment you want a big model with room to breathe, 16GB is the wall you meet.

Where models fit, the speed is the point. Expect something like 70 to 100 tokens/sec on a 13B at Q4, among the quickest you’ll see on a single consumer card, and it holds that pace through long context and heavier prompts where slower cards start to sag. On value, roughly £1000 for 16GB is about £63 per gigabyte, the steepest £/GB-VRAM figure here by a distance; you’re paying for pace and a warranty, full stop. Software is current Blackwell CUDA, well supported across the major inference stacks, though a brand-new architecture occasionally wants an up-to-date build of your tools before everything clicks into place.

Who should buy it, and who shouldn’t

Buy it if you want the fastest new 16GB card, you value a warranty and current drivers, and your models really do live under 16GB. For quick 7B to 14B work it’s superb, and it doubles happily as a top-tier gaming card when you’re not running models. If new and fast is the brief, this delivers.

Skip it if VRAM is your priority; a used 3090 or 4090 gives you 24GB and runs models the 5080 simply can’t hold, and for big-model work capacity beats speed every time. If you want new and you can stretch further, the 5090’s 32GB is the far more future-proof buy. But for pure single-card speed on models that fit, the 5080 is hard to beat.