Modelling and rendering want different machines
Most practices buy one specification for everybody. Usually the expensive one, because the visualiser asked loudest. Eight architects then end up with graphics cards doing very little.
This has always been wasteful. In 2026 it is expensive. AEC Magazine reported in February that a 96GB DDR5 kit had gone from around £200 to around £800 and 256GB of ECC memory from about £1,500 to about £4,000, with shortages expected to run into 2027. Memory is now often the largest single line in a workstation quote, which makes buying the same amount of it for everyone an increasingly odd decision.
The fix is not a better machine. It is two or three specifications instead of one.
The two jobs pull in opposite directions
Modelling and documentation lean heavily on one core running fast. Dragging a wall, regenerating a view, opening a schedule: most of that work happens in a single thread, so clock speed matters far more than core count. A sixteen-core processor does not draw a section any quicker than an eight-core one with a higher boost clock.
Rendering does the opposite. It spreads across everything available, whether that is every core in the processor or the GPU, depending on which engine the practice uses. Here core count and graphics hardware pay for themselves, because the job is genuinely parallel.
Buy one machine for both and you overspend on cores and graphics for the people who model, while the people who render still wait.
What each group actually needs
- Modelling and documentation. Highest single-core clock you can get, memory sized to the largest model they open rather than to a round number, fast local NVMe storage, a mid-range GPU. This is most of the studio
- Visualisation and analysis. Core count or GPU according to the renderer, roughly double the memory and a machine that can be left working overnight without stopping someone else
- Everyone. Enough memory for the model as it is now, not as it was when the machine was bought. This is the single most common cause of a "slow computer" ticket
Work out who is who before you buy
The split is usually obvious once someone asks. Not from job titles, which mislead, but from what people actually open. A part two who renders twice a year is a modelling machine. A technician who runs overnight batches is not, whatever their title says.
Ask three questions across the studio: what software is open at 3pm on a normal Tuesday, how large is the biggest thing you open and when did you last wait more than a minute for the machine rather than the model. Half an hour of that is worth more than any benchmark chart.
What it saves
A practice of twenty with one specification typically buys twenty visualisation machines and uses three of them that way. Split into fifteen modelling and five rendering and the modelling machines get a faster clock than the single spec had, the rendering machines get more memory than they had and the total comes down. The people who wait stop waiting. The people who never waited never notice.
It also makes the refresh cycle sane. Modelling machines age gracefully and can be handed down. Rendering machines earn a shorter cycle because the work they do keeps growing.
What would you find if you asked everyone what they had open at 3pm yesterday? We spec, supply and look after CAD and AutoCAD workstations, the Revit and BIM environment around them and the everyday IT support London architecture practices run on. If a refresh is coming, start with the numbers or tell us what your studio opens.
