The question this piece answers: how many GPUs can a home server actually feed, and which specification decides it? Not the slot count, and not the power supply. It is the lane budget — the number of PCIe links the processor exposes, how the board is allowed to split them, and what the chipset's own uplink costs you. The platform numbers below come from AMD's and Intel's own product pages and press materials, re-checked on 2026-11-17; the bandwidth figures are derived from the PCI-SIG encoding rules and shown as arithmetic so you can check them; every other number marked "ours" is a calculation, not a vendor claim.

What is a PCIe lane budget?

A PCIe lane is one bidirectional serial link, and devices are wired in groups of 1, 2, 4, 8 or 16 of them. A "x16 slot" is sixteen lanes bonded into one link. The lane budget is how many lanes the CPU can supply in total — a hard, physical limit set by the package and the socket, not by how many slots the board vendor soldered on. Every device you attach spends from that budget: a GPU at x16 spends 16, an NVMe drive typically spends 4, a 25 GbE NIC spends 4, and the chipset itself spends 4 on a desktop platform.

The bandwidth per lane is fixed by generation and encoding. PCIe 4.0 runs at 16 gigatransfers per second and uses 128b/130b line coding, so one lane carries 16 × 128 ÷ 130 ÷ 8 = 1.97 GB/s; PCIe 5.0 doubles the transfer rate to 32 GT/s and carries 3.94 GB/s per lane. Multiply out and the familiar numbers appear: x16 at 4.0 is 31.5 GB/s, x16 at 5.0 is 63.0 GB/s, and x4 at 5.0 (15.75 GB/s) equals x8 at 4.0. That equivalence is the one to keep in mind, because it is how a board vendor turns four lanes into a "PCIe 5.0 slot".

What does each platform actually give you?

Here is the published budget, by platform, from the manufacturers' own specifications.

PlatformPCIe budgetPractical GPU countWhat eats the rest
AM5 desktop (Ryzen 9000)28 native / 24 usable PCIe 5.01 at x16, or 2 at x8/x8 with storage moved4 lanes to the chipset, 4 to the primary M.2; X870E adds 8 Gen4, X670E adds 12 Gen4 from the board
Intel Xeon 6 6900P96 PCIe 5.0 (or 64 CXL 2.0)4 at x16, or 8 at x8NVMe, NICs, and any CXL devices share the same 96
AMD EPYC 9005 (SP5)128 PCIe 5.0 per socket, 160 dual-socket8 at x16 single-socket12 memory channels are separate; lanes go to GPUs, NVMe and fabric NICs
Threadripper PRO 7000 WX + WRX90148 total / 144 usable, up to 128 at Gen54–7 depending on bifurcation8-channel memory; TRX50 boards drop to 92/88 with 48 at Gen5
EPYC 8005 (SP6)96 PCIe 5.0 + 8 PCIe 3.04 at x166 memory channels; the edge-focused socket

Two details in that table are worth more than the headline numbers. The first is the AM5 chipset uplink. AMD's own spec pages list "Native PCIe Lanes (Total/Usable) 28, 24" — the four missing lanes are the link to the chipset, which then fans out to SATA, USB, Ethernet and the secondary slots. Everything hanging off the chipset shares those four lanes, which is why a board with "three x16 slots" will run the third one at x4 through the chipset, and why an NVMe drive in a chipset-attached M.2 slot can contend with your network card. The second is bifurcation. A x16 slot can be split into x8/x8 or x4/x4/x4/x4, but only if the board is wired and the firmware allows it — and splitting is how a 24-lane platform pretends to host two cards at x8 each while quietly moving storage to the chipset.

Does x8 versus x16 change tokens per second?

Almost not at all, and this is where most homelab GPU advice goes wrong. During decode — the token-by-token generation that dominates wall-clock time for a chat workload — the weights are resident in VRAM and the model reads them from HBM, not from the host. The PCIe link carries the prompt in once and the tokens out once per request. A single-GPU inference server therefore runs at essentially the same speed on x4 as on x16; the bus is idle most of the time. What the link does set is how long the model takes to get into VRAM in the first place.

LinkBandwidth40 GB of weights loads inFour GPUs, sequential
PCIe 4.0 x47.88 GB/s5.1 s20.4 s
PCIe 4.0 x815.75 GB/s2.5 s10.2 s
PCIe 4.0 x1631.5 GB/s1.27 s5.1 s
PCIe 5.0 x1663.0 GB/s0.63 s2.5 s

Those figures assume the storage can keep up and the loader is not the bottleneck, which is why they are an upper bound on the link's contribution. The practical reading (ours): if your workflow restarts a 40 GB model twenty times a day, the difference between x4 and x16 is about 77 seconds a day — real but small. If it restarts a hundred times a day, or if you are loading a 400 GB mixture-of-experts checkpoint, the same ratio becomes minutes per hour, and the lane budget starts to matter more than the GPU's clock speed.

When does the lane budget actually bite?

Four cases, in descending order of how often they surprise people.

  • Adding the second GPU. On AM5 the second card usually costs you x8/x8 — acceptable for inference, but it also forces storage onto the chipset's four lanes, which is where the "why is my dataset load slow" complaints come from. On a 96-lane Xeon 6 or 128-lane EPYC, a second and third card cost nothing that matters.
  • Multi-GPU without NVLink. Consumer and most workstation cards have no NVLink, so tensor parallelism runs its collectives over PCIe. The traffic per token is small — tens of kilobytes per layer — but the latency per collective is not free, and with dozens of layers and two collectives per layer the round trips add up. Narrow links hurt here through latency, not bandwidth.
  • Checkpointing and dataset streaming. Training writes state to disk constantly; that traffic shares the same lanes as the GPUs if the NVMe is CPU-attached. This is the case where x4 really is a mistake, and where a platform with 96+ lanes pays for itself without a single extra FLOP.
  • Host-to-device copies for offload. Any workflow that streams layers or experts from system RAM to VRAM is bounded by the link — the one inference pattern where lane width becomes a throughput number rather than a loading-time number.

What should a home builder actually buy?

The decision collapses to a single question: do you intend to run more than one accelerator? If the answer is no, an AM5 board with 24 usable lanes is not a compromise — a single GPU at x8 loses nothing measurable in tokens per second, and the money is better spent on VRAM. If the answer is yes, then the lane budget is the first specification you shop, before the GPU: a used EPYC or a Threadripper platform with 96–128 lanes costs more than a desktop board and less than the second GPU it makes useful.

Two buying rules follow (ours). Count lanes, then slots. Add up what you plan to attach — 16 per GPU, 4 per NVMe, 4 per NIC — and compare against the usable figure, not the native one; the 4 lanes the chipset took never come back. Check the board's bifurcation table before the CPU. A 128-lane processor on a board that only wires one x16 slot and three x4 slots is a 128-lane processor you cannot use, and the manual's slot diagram is the only place that is written down.

What are the caveats?

Lane counts are per socket. Dual-socket EPYC reaching 160 lanes is a property of the pair of processors and the board's wiring, not something a single CPU delivers; a 2P board also splits memory across sockets, which changes NUMA behaviour for anything you pin to a GPU. Bandwidth figures are theoretical. Real transfers lose a few percent to protocol overhead, and a 40 GB load also has to be read from disk, so the link is rarely the only bottleneck. The x4-is-fine conclusion is workload-specific. It holds for single-GPU inference with the model resident in VRAM; it does not hold for training, for offload-heavy serving, or for multi-GPU tensor parallelism at large batch. And none of this replaces the power and thermal budget. Four GPUs at x16 is a ~2.5–3 kW machine before you count the rest, which is a different article in this series — the branch circuit and the room, not the bus.