Processors

In a GPU machine the CPU decides something other than you expect

In an ordinary computer the processor does the maths. In a multi-GPU station it mostly feeds the cards with data. So the numbers to check are not the advertised clock speed but PCIe lane count, supported memory capacity and the number of memory channels.

A quick sanity check. Each card wants sixteen lanes. Four cards means sixty-four lanes plus lanes for drives and the network adapter. A desktop CPU offers about twenty. After that lanes get split through a switch and the cards start waiting for each other.

What to look at in the spec

• PCIe lanes and generation. Two cards on PCIe 5.0 x8 perform like 4.0 x16. That saves a build when lanes are short.
• Memory channels. Eight or twelve channels deliver bandwidth a four-channel platform simply cannot reach.
• Cores versus clock. Data preparation and virtualisation want cores. CAD and single-threaded work want clock speed.
• Socket and thermal envelope. Three hundred watts in a case without planned airflow means throttling under load.

Typical pairings

One or two cards in a workstation: an HEDT-class CPU with enough lanes will do. Four cards or more, storage and fast networking: a server platform where lanes and memory channels come with headroom.

Tell us how many cards you plan and what they will run. We will point out where the CPU becomes the bottleneck and where paying more buys nothing.