Hardware / two extremes
From a dual-GPU AI workstation made to run heavy local AI, to a computer smaller than a credit card — built, not bought.
I wanted to run large AI models locally — the kind that power Raven — without paying cloud bills forever. The trick is VRAM. So I built an AM4 machine that pairs a normal gaming GPU with a second-hand NVIDIA Tesla P40, a datacentre card that gives 24GB of VRAM for a fraction of the price of a modern equivalent. The result is a quiet, single-tower workstation with serious memory for AI, built mostly from carefully-chosen used parts.
Raven runs many things at once — scraping, processing, models and automation in parallel. More cores means none of it has to wait its turn.
orchestrating Raven's workloads and crunching large batches of scraped business data at speed.
Big datasets and several model contexts need to sit in memory together, so nothing grinds to a halt halfway through a run.
holding large scraped datasets and multiple live tasks in memory at the same time.
VRAM is the hard ceiling on which AI models you can run locally — more of it means bigger, smarter models actually fit on the card.
running the large local language models that power Raven, fully offline and free.
Model files are gigabytes each and scraped data piles up fast; the NVMe loads models instantly, the big SATA drive holds everything else.
keeping a whole library of models plus datasets and project files on hand.
A modern 24GB card costs more than this whole machine. The P40 gives the same VRAM for AI work for around £200 — the single decision the build is designed around.
AM4 is mature and cheap on the used market, yet a 16-core 5950X is still seriously fast. More cores and VRAM per pound than chasing the newest platform.
Everything runs on my own machine — no API keys, no usage limits, no monthly bill. It is the same principle that makes Raven free to run forever.
At the other extreme: a full Linux computer built around a Raspberry Pi Zero 2 W and a microSD card, for under £50 all-in. A quad-core machine with wireless built in, small enough to disappear into a project. Proof that the interesting part isn't the budget — it's knowing what to do with it at either end of the scale.