The practical rule
Add up what your largest open project takes (the software shows it) and multiply by two or three: the program, the system, the browser and temporary copies live there too. If the figure exceeds the machine’s memory, the system starts using the disk and everything slows down without anything failing. Better to keep a third free on a normal day.
How much by type of work
| Task | Memory | Workstation | Price |
|---|---|---|---|
| CAD of parts and medium assemblies | 32 to 64 GB | Creative Pro, 64 GB | €5,280 |
| BIM with federated models, large-scene rendering, 8K video | 64 to 128 GB | Pro, 64 GB ECC | €6,770 |
| Large projects with room to grow | 128 GB | Advanced Plus, 128 GB registered ECC | €14,265 |
| Simulation, point clouds, data science | 128 to 256 GB | Enterprise, 256 GB ECC in eight channels | €30,435 |
| Very large models, genome assembly, datasets that do not fit | 512 GB or more | Enterprise Supreme, 512 GB ECC | €57,400 |
For artificial intelligence the GPU memory rules; system memory should be at least double, to load and prepare the data.
Channels matter as much as gigabytes
Memory is read through channels: two on a Ryzen, four on a Threadripper, eight on a Threadripper PRO and up to twelve on an EPYC. More channels, more bandwidth, and some tasks (simulation, in-memory databases, data science) run at the speed of that bandwidth, not of the processor. That is why it pays to fill every channel with identical modules, even if smaller, rather than fitting two large ones.
When ECC memory is worth it
ECC memory detects and corrects single-bit errors on the fly. Such an error on an office PC is never noticed; in an eight-hour render or a computation that lasts days it shows up as a wrong result and nobody knows why. It pays when the machine works for hours unattended, when it handles data that cannot be repeated and on any Threadripper or EPYC, which use it as standard in its registered version (RDIMM). On a Ryzen it exists in an unbuffered version (ECC UDIMM) with motherboards that support it.
Buying today with expansion in mind
Leave free slots: a Threadripper PRO has eight and a Threadripper four, so with half filled you can double the memory without discarding anything. And buy modules of the same type and speed: mixing is the quickest way to an unstable machine. Our workstations ship with every channel filled and the upgrade path stated on the product page.
Related hardware and guides
Frequently asked questions
Is 64 GB better in two modules or four?
On a Threadripper, in four: that way all four channels are used and bandwidth doubles. On a Ryzen, with two channels, two modules are enough.
Is ECC memory slower?
Registered memory (RDIMM) has slightly higher latency, imperceptible in real work. In return it corrects errors and allows capacities that normal memory cannot reach. In long computations, reliability is worth more than that margin.
Can I fit ECC memory on a Ryzen?
Yes, unbuffered ECC (UDIMM) on motherboards that support it; our Pro workstation carries it as standard. Registered memory (RDIMM) is only for Threadripper, EPYC and Xeon.
How much memory do I need for artificial intelligence?
GPU memory rules: the model has to fit in it. System memory should be at least double the GPU’s to prepare data and load models. For training, more.
Shall we size your workstation?
Tell us which software you use and the size of your projects. We will propose the processor, memory, GPU and drives, with a price, within 12 working hours.