If you want a new NVIDIA graphics card for running local AI models, such as chatbots in Ollama or LM Studio, as well as gaming, the GeForce RTX 5070 Ti and RTX 5080 are the obvious 16GB choices below the very expensive RTX 5090. The 5080 is faster on paper, but for local AI the more important question is what each card can actually hold, and there the answer is identical.
Forge PC has not benchmarked these cards for AI. This comparison uses NVIDIA's official UK specifications, memory bandwidth figures reported by TechRadar and NVIDIA's own technical explanation of how language models run on a GPU. UK prices are approximate, from SpecPrice's cheapest-GPU tracker (updated 15 November 2026, 01:22 GMT), and will change.
Specifications compared
| RTX 5070 Ti | RTX 5080 |
|---|
CUDA cores | 8,960 | 10,752 |
AI TOPS (NVIDIA figure) | 1,406 | 1,801 |
Memory | 16GB GDDR7 | 16GB GDDR7 |
Memory bus | 256-bit | 256-bit |
Memory bandwidth (TechRadar) | 896GB/s | 960GB/s |
Boost clock | 2.45GHz | 2.62GHz |
Total graphics power | 300W | 360W |
NVIDIA's recommended PSU | 750W | 850W |
Cheapest UK price (15 November 2026) | about £1,268 (Box) | about £1,520 (Scan) |
Same 16GB, same models
The amount of graphics memory decides which models you can run at full speed, because the model's weights need to fit on the card. Both cards have 16GB, so any model and quantisation that fits on one fits on the other. Neither card lets you step up to a bigger model than the other. Our guide to how much VRAM you need and our explainer on LLM quantisation and GGUF files show what fits in 16GB.
Why bandwidth matters most for chat
NVIDIA's technical blog on LLM inference explains that a language model works in two phases. The prefill phase processes your prompt in parallel and "effectively saturates GPU utilization", so it benefits from more compute. The decode phase, where the model writes its answer one token at a time, is memory-bound: NVIDIA says model execution is frequently bound by memory bandwidth, "in particular, bandwidth-bound in the weights".
That matters here because the two cards' memory bandwidth is close. TechRadar reports 960GB/s for the RTX 5080 against 896GB/s for the RTX 5070 Ti, about 7% more. The 5080's compute advantage is bigger: NVIDIA lists 1,801 AI TOPS against 1,406, about 28% more, and 20% more CUDA cores.
In practice, that suggests the 5080 should process long prompts and documents faster, while the speed at which it writes answers should be only modestly ahead, because that part is limited mainly by bandwidth. We have not measured this ourselves, so treat it as what the specifications imply rather than a benchmark result.
What about image generation?
We have not found an independent, like-for-like image generation comparison of these two cards to quote. The 5080's extra CUDA cores and AI TOPS may help there, but we cannot put a number on it. The memory limit is the same 16GB on both cards, so the image models you can load do not change.
The price gap
On SpecPrice's tracker, updated at 01:22 GMT on 15 November 2026, the cheapest RTX 5070 Ti was about £1,267.99 at Box and the cheapest RTX 5080 about £1,519.99 at Scan, a gap of about £252. For a card that holds exactly the same models and has only about 7% more memory bandwidth, that is a big premium for local AI use.
It is also worth knowing what else is around that money. The same tracker listed AMD's Radeon AI PRO R9700, which AMD specifies with 32GB of GDDR6 and 640GB/s of bandwidth, at about £1,609.99 at Scan. It has less bandwidth than either GeForce card but twice the memory, so it can load much larger models. Our NVIDIA vs AMD for local AI guide covers the software trade-offs, and our used GPU guide looks at second-hand options.
Which should you buy?
Buy the RTX 5070 Ti if local AI is a big part of why you are buying. It runs the same models as the 5080 for about £250 less (15 November 2026), with only slightly less memory bandwidth.
Buy the RTX 5080 if gaming performance is your priority, or you often feed long documents to models, where its extra compute should help.
Consider a 32GB card or two GPUs if you want to run larger models than 16GB allows. Our guide to running a local LLM on two GPUs explains how that works.
For the wider picture, see our best GPU for local LLMs guide, and for gaming value, RTX 5080: buy or wait?.
Our verdict
For local AI, the RTX 5070 Ti is the better buy. Both cards have 16GB of GDDR7 on a 256-bit bus, so they run exactly the same models, and TechRadar's figures give the RTX 5080 only about 7% more memory bandwidth, which NVIDIA says is what usually limits how fast a model writes. The 5080's roughly 28% higher AI TOPS should help with long prompts, but at about £252 more on 15 November 2026 it is hard to justify for AI alone. Forge PC has not benchmarked these cards; the comparison uses NVIDIA and AMD specifications and TechRadar's bandwidth figures.