Which tasks scale with several GPUs
GPU rendering (Octane, Redshift, V-Ray GPU, Blender Cycles) scales almost linearly: two cards, half the time. Inference of large models that do not fit in one card is split across several, and training of medium models uses two or four. And when several people share the machine or several jobs run at once, each GPU serves one.
Which tasks do not scale
CAD, BIM, video editing and real-time visualisation use one GPU; a second one speeds up nothing. Nor do small AI models that fit in one card: they run faster on one good card than on two medium ones. And the memory of two GPUs does not add up for a single model unless the software splits it explicitly.
The motherboard: PCIe lanes and spacing between slots
Each GPU wants a PCIe x16 slot with enough lanes. A Ryzen has lanes for one GPU at x16 or two at x8; a Threadripper (TRX50) for three at x16; a Threadripper PRO (WRX90) has seven slots and lanes for four two-slot GPUs or more. You also need spacing between slots: a three-slot card covers the next one; two-slot blower RTX PRO cards are designed to sit side by side.
Power supply, case and cooling
Two RTX 5090s need a 2,000 W power supply and a large case with fans front and rear; three or four RTX PRO 6000 Max-Q, at 300 W each, fit with a 1,600 to 2,000 W supply and exhaust their air out the back through the blower, without overheating the neighbour. For machines that work 24 hours a day, a server case with front-to-back airflow and, if needed, rack mounting.
| Workstation | GPUs | Power supply | Price |
|---|---|---|---|
| Enterprise: Threadripper PRO, 256 GB ECC, server case | 1 × RTX 5090, expandable up to 7 slots | 1,600 W | €30,435 |
| Enterprise Plus: 32-core Threadripper PRO | 2 × RTX 5090 | 2,000 W | €38,200 |
| Enterprise Pro: 64-core Threadripper PRO | 2 × RTX 5090 | 2,000 W | €42,605 |
| Enterprise Supreme: 512 GB ECC | 2 × RTX 5090, expandable with RTX PRO | 2,000 W | €57,400 |
In the configurator the RTX 5090s can be swapped for RTX PRO 4500, 5000 or 6000 Max-Q, and a third and fourth added.
What nobody mentions: noise, heat and electricity
Two GPUs at full load are more than a kilowatt of heat in the room and a machine you can hear. If it will work at night or in an office, better in a technical room or a rack, and on a socket that can take it. Our multi-GPU machines pass the 24-hour test with every GPU loaded, which is exactly what they will do in your office.
Related hardware and guides
Frequently asked questions
Do two GPUs add up their memory?
Not on their own. For rendering, each GPU needs the whole scene in its memory. For artificial intelligence, the software can split a model across several cards, but it has to be configured and loses some speed.
Can I fit two GPUs on a Ryzen?
Yes, at x8 each on a motherboard that supports it, like our Pro workstation. For three or more, you need a Threadripper or a Threadripper PRO for the PCIe lanes and the space.
GeForce or RTX PRO for several GPUs?
For two, GeForce is cheaper. For three or four, the 300 W blower RTX PRO cards: they sit side by side, do not overheat each other and need a smaller power supply. Three-slot GeForce cards cover the next slot.
What power supply do I need?
Two RTX 5090s, 2,000 W. Two RTX PRO 6000 Max-Q, 1,600 W. Four Max-Q, 2,000 W. And a socket and a UPS sized for it.
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.