A budget B550 board cannot host two GPUs for compute rental

The symptom

Two GPUs installed: no network, and the machine could not be listed for rental. One GPU removed: everything worked immediately. The same pair of cards ran fine on an X570 board.

Two things going wrong at once

First, address space. Modern GPUs expose a large 64-bit memory window through Resizable BAR. With two of them and Above 4G Decoding disabled, the firmware runs out of space below 4 GB and devices fail to initialise — including the onboard network controller, which is a PCIe device like any other. That is why the LAN dies rather than just the second GPU.

Second, topology. On B550 the first slot takes all 16 CPU lanes. The second full-length slot is wired to the chipset — typically PCIe 3.0 x4, sometimes x2, and it often shares lanes with the second M.2. That link is also the path your network controller uses.

Why X570 behaves differently

On a good X570 board the top two slots are wired as CPU x8/x8. The second GPU gets CPU lanes and the chipset — with the NIC hanging off it — is left alone. That is the whole difference.

Enabling Above 4G Decoding, turning on Resizable BAR and booting pure UEFI may get a B550 to POST with two cards. But the second GPU still sits on a contended chipset link shared with the network interface, which is a poor foundation for a machine whose income depends on staying connected.

What to build

Single-GPU machines: any decent B550 is fine and completely reliable.

Two or more GPUs: use a platform that gives the second card CPU lanes — X570 on AM4, or B650/X670 on AM5. Do not pair a multi-GPU board with low-VRAM cards; the platform overhead only makes sense with cards that earn enough to justify it.

For rental workloads specifically, PCIe bandwidth barely matters once the model is loaded — each GPU runs its own job. What matters is that the machine stays online.