How Birmingham AV builds an AI ready PC

The phrase "AI ready" gets stuck on a lot of hardware that is not, in fact, ready. A modest office tower with 16GB of RAM and integrated graphics will run a hosted chatbot in a browser tab, but it will not run a 13 billion parameter model locally, and it will crawl the moment a Stable Diffusion checkpoint is loaded. This piece sets out the specification Birmingham AV recommends for local inference, agents, and image models.

The four numbers that matter

Ignore marketing bullet points. Look at four numbers on the spec sheet.

• RAM: 32GB minimum. 64GB is the sweet spot for anyone loading larger context windows or running multiple containers.
• VRAM: 12GB minimum on the GPU. Below this figure you are excluded from most useful open weight models at usable quantisation levels.
• Storage: 2TB NVMe minimum, PCIe Gen4 or Gen5. Model weights are large and slow storage bottlenecks everything.
• CPU: a modern desktop chip with at least 8 cores and PCIe Gen4 support, so the GPU is fed properly.

A quality power supply (650W Gold rated or better) and adequate cooling round the build out. Skimp on the PSU and you will chase instability.

Why 12GB VRAM is the real floor

Video memory is the hardest constraint in local AI work. A 7 billion parameter model at 4 bit quantisation lands around 4 to 5GB of VRAM, but the moment you increase context length or add image generation alongside chat, you eat headroom fast. 8GB cards force compromises. 12GB opens up 13B class models comfortably. 16GB is where SDXL, Flux, and 20B class models sit.

The RTX 5070 Ti with 16GB of GDDR7 is the honest inflection point between hobbyist and productive for local AI. It handles inference on Llama 3 class 13B models, SDXL image generation, and light fine tuning, all without hitting the memory wall that plagues smaller cards.

The BAV recommended build

For most buyers the answer is the Ryzen 7 5700X paired with an RTX 5070 Ti.

• Ryzen 7 5700X: 8 cores, 16 threads, 65W TDP, PCIe Gen4. Feeds a modern GPU without bottlenecking, and sits on the mature AM4 platform where DDR4 memory is cheap.
• RTX 5070 Ti 16GB: best value tier for VRAM per pound in the Blackwell generation, with strong support in PyTorch, llama.cpp, ComfyUI, and Ollama.
• 32GB DDR4 3600MHz (upgradeable to 64GB on the same board).
• 2TB NVMe Gen4 SSD, typically a Samsung 990 Pro or WD SN850X class drive.
• 750W 80 Plus Gold PSU from Corsair, Seasonic, or be quiet!
• Mid tower case with front mesh and at least three intake fans.

Complete builds in this configuration are available in the Birmingham AV eBay store. Listings span single 2TB NVMe through dual drive setups with additional 4TB SATA storage, and memory configurations from 32GB up to 64GB, so buyers can pick the tier that matches their workload.

Why not go bigger

The temptation is to reach for a Ryzen 9 or an RTX 5080. Both are excellent chips. Neither is necessary for the workloads described here, and both push total build cost past the point where a used RTX 3090 with 24GB becomes the more sensible choice for serious model work.

The 5700X plus 5070 Ti build lands in a price band where the machine is genuinely useful for local AI and remains a strong gaming rig. If workloads grow, the GPU can be swapped without replacing the whole system, because the PCIe Gen4 lanes and PSU headroom are already there.

Ubuntu or Windows 11 Pro

Both operating systems run modern AI stacks well. The choice is about workflow.

• Ubuntu 24.04 LTS: preferred for serious inference server work, Docker containers, or deployment onto Linux servers. CUDA drivers are first class, and headless operation is straightforward.
• Windows 11 Pro: preferred for buyers who also want a general purpose workstation. WSL2 runs CUDA workloads directly on the host GPU, so Linux tooling is available without a dual boot.

Birmingham AV ships either, or dual boot on request. Windows 11 Pro is the default on prebuilt listings.

Cooling, noise, and longevity

AI workloads pin the GPU at full load for long periods. This is different to gaming, where thermal peaks are short. A card that runs cool during a two hour raid will run hot during an eight hour training run, and heat kills components over time.

Every AI ready build from Birmingham AV includes tower CPU cooling (not the stock cooler), at least three case fans in a balanced configuration, and a mesh front case. Fan curves are tuned for sustained load. This is the kind of detail that does not appear on a spec sheet but decides whether the machine is still healthy in three years.

FAQ

Can I run local AI on a machine with 8GB VRAM?

Yes, with real limitations. Small models (7B parameters at 4 bit quantisation) will run, and basic Stable Diffusion 1.5 image generation is possible. You will not run SDXL comfortably or 13B class models with long context, and you will hit out of memory errors regularly. It is a starting point, not a destination.

Is AMD Radeon a viable alternative to NVIDIA for AI work?

Improving, but not equivalent yet. ROCm support has matured and llama.cpp runs well on RDNA3 and RDNA4. However, PyTorch, ComfyUI, and most fine tuning tooling still assume CUDA. For a three year horizon, NVIDIA remains the safer choice for AI specifically.

How much does 64GB of RAM actually help?

More than most people expect. Loading large models into system RAM before they move to VRAM, running Docker containers with vector databases, and keeping browsers plus IDE plus local inference server resident all become smoother. If the budget stretches, 64GB is the best incremental upgrade on this build.

Will this build handle video generation?

For short clips at modest resolution, yes. AnimateDiff and current open weight video models run on 16GB of VRAM. Longer clips and higher resolutions push into 24GB territory, at which point a used RTX 3090 or a step up to the RTX 5080 becomes relevant.

About Birmingham AV

We have sold over 87,000 items on eBay since 2017, with 24,756 buyer feedbacks at 98.9% positive. Every system ships with a twelve month warranty. Birmingham AV Ltd is registered at Companies House under number 12383651, VAT registration GB 348755066, based in Bromsgrove, Worcestershire. One of the highest volume refurbished PC operations on eBay UK, we build to a specification we would use ourselves.