If you want to run large language models (LLMs) on your own PC, video memory is the first thing to check, because the model has to fit in it to run quickly. Only a few cards offer 32GB: NVIDIA's flagship GeForce RTX 5090 and AMD's workstation-focused Radeon AI PRO R9700. They share a memory size but almost nothing else, including price.
Forge PC hasn't tested either card. Specifications come from AMD and NVIDIA, performance figures are from a published llama.cpp community comparison and Puget Systems' testing (credited below), and prices are approximate UK prices from SpecPrice on 16 November 2026.
Specifications
| Radeon AI PRO R9700 | GeForce RTX 5090 |
|---|
Architecture | AMD RDNA 4 | NVIDIA Blackwell |
Video memory | 32GB GDDR6 | 32GB GDDR7 |
Memory bus | 256-bit | 512-bit |
Peak memory bandwidth | 640GB/s (AMD) | Not listed on NVIDIA's UK spec page |
Cores | 64 compute units, 4,096 stream processors, 128 AI accelerators | 21,760 CUDA cores |
Boost clock | Up to 2,920MHz | 2.41GHz |
Board power | 300W | 575W |
Recommended PSU | 750W minimum | 1,000W system power |
Interface | PCIe 5.0 x16 | PCIe 5.0 |
The biggest practical difference for LLMs is memory bandwidth. Generating each word (token) means reading the model's weights from video memory, so a wider, faster memory bus usually means faster replies. The RTX 5090 has double the bus width and faster GDDR7 memory; the llama.cpp comparison below puts its bandwidth at about 1.8TB/s.
What published tests show
In a llama.cpp discussion on GitHub from February 2026, one user ran the same model file, Qwen3.5-35B-A3B at Q4_K_XL, on both cards, using the same llama.cpp commit and settings. The RTX 5090 used NVIDIA's CUDA backend on a Ryzen 9 9900X system; the R9700 used the Vulkan backend on a six-core Ryzen 5 7500F system, so the PCs weren't identical.
Generating text: the RTX 5090 averaged 194.0 tokens per second, against 127.4 for the R9700, making it about 1.5 times faster.
Reading prompts: the RTX 5090 was 2.6 times faster with a 512-token prompt, rising to 3.4 times faster at 32,768 tokens.
That second figure matters if you paste in long documents or use coding assistants that resend large amounts of context. The tester also noted that 127 tokens per second already felt very comfortable for chat, and that this model is a mixture-of-experts (MoE) design, which reads fewer weights per token than a dense model of similar size. Gaps on dense models may be larger.
Puget Systems tested the R9700 with vLLM, including a dual-card setup on a bare-metal Linux host. With two cards sharing a 27-billion-parameter Qwen model in FP8, it measured about 15.9 tokens per second for a single user and about 156 tokens per second in total across eight simultaneous users with stock vLLM. These are Puget's results, not ours, and speeds vary widely with the model, quantisation and software.
Software support
NVIDIA's CUDA is still the default target for most local AI tools, so the RTX 5090 tends to work with the least effort. The R9700 relies on AMD's ROCm or the cross-vendor Vulkan backend. In the llama.cpp discussion, the tester, who ran the R9700 on Windows, said the Vulkan backend needed no ROCm installation and works out of the box on both Windows and Linux, while another user reported that switching to ROCm on Ubuntu took more work. Our guide to running local AI on an AMD Radeon GPU covers the main options.
UK prices and the two-card option
On 16 November 2026, SpecPrice listed:
the cheapest Radeon AI PRO R9700 at about £1,610 (a PowerColor card at Scan, the only UK shop it tracked for this model)
the cheapest RTX 5090 at about £7,000 (a Gigabyte AORUS XTREME WATERFORCE at Scan), with the next listing at Box about £1,450 higher
That makes the RTX 5090 roughly 4.3 times the price. Two R9700s, about £3,220, would give you 64GB of video memory for less than half the price of one RTX 5090, although you'd need a motherboard with two suitable slots, a bigger power supply and software that can split a model across cards. Our guide to running an LLM on two GPUs explains how that works.
Gigabyte offers exactly this choice in its AI TOP 100 B850 desktop. It says the RTX 5090 version generates up to 64.9 tokens per second on a 32-billion-parameter model, while the dual-R9700 version can run models of up to 235 billion parameters. Both are Gigabyte's internal figures.
The PowerColor AI PRO R9700 32GB is in our specs database.
Which should you buy?
Choose the Radeon AI PRO R9700 if: your main aim is fitting 32GB models (or 64GB with two cards) for the lowest price, you're happy with Vulkan or ROCm, and chat-speed generation is fast enough for you. Its 300W board power is also much easier to cool and power.
Choose the RTX 5090 if: you need the fastest prompt processing for long documents and coding agents, want the widest software support through CUDA, or also want a top-end gaming card. Make sure your power supply meets NVIDIA's 1,000W recommendation; see our PSU guide.
For a wider look at the trade-offs, read NVIDIA vs AMD for local AI and our best GPU for local LLMs.
Our verdict
The RTX 5090 is clearly the faster card: in a published llama.cpp comparison it generated text about 1.5 times faster than the Radeon AI PRO R9700 and read long prompts up to 3.4 times faster, with easier CUDA software support. But both have 32GB, and at about £1,610 against about £7,000 on 16 November 2026 the R9700 is far better value for fitting large models, while two R9700s give 64GB for less than half the price of one RTX 5090. Forge PC hasn't tested either card.