ATX vs mATX for Multiple GPUs

ATX vs mATX for Multiple GPUs | Generated by AI

Home 2026.07

Question: What is a PCIe extension cable in the context of using multiple GPUs, and why do people recommend ATX motherboards instead of mATX for two or more GPUs?

Answer:

A PCIe 延长线 (PCIe riser cable / PCIe extension cable) is a cable that extends the PCIe slot so you can move the GPU away from the motherboard.

Normally:

GPU
 |
 |  (direct connection)
 |
PCIe x16 slot on motherboard

With a riser:

Motherboard PCIe slot
        |
        |
   PCIe extension cable
        |
        |
      GPU

The GPU no longer has to physically sit directly above the motherboard slot.

Common examples:

Mining rigs
AI multi-GPU servers
Small cases
Workstations with large GPUs

A typical riser has:

PCIe x16 motherboard adapter
        |
        |
 ribbon cable
        |
        |
PCIe x16 GPU adapter

Usually the electrical connection is still only PCIe x1/x4/x8, depending on the riser. Many mining risers use PCIe x1 because bandwidth is not important for mining.

For AI inference/training, you generally want:

PCIe x16 physical slot
PCIe x8 electrical minimum per GPU if possible
high quality shielded riser if using extension

Why ATX instead of mATX for 2+ GPUs?

The problem is not only “number of slots”; it is:

Physical spacing
PCIe lanes
Cooling

Example:

mATX board

CPU

PCIe x16
========
RTX 3090 (3-slot)

PCIe x4
====
blocked

A RTX 3090 / 4090 is usually 2.5~4 slots thick.

After installing one GPU:

GPU
GPU cooler
GPU cooler
GPU cooler

it covers the next PCIe slot.

So even if the motherboard has:

PCIe x16
PCIe x16

physically there is no room.

ATX:

CPU

PCIe x16
========
(empty spacing)

PCIe x16
========

PCIe x4
====

Example:

RTX3090
  |
  | 2 slots gap
  |
RTX3090

Both cards can breathe.

PCIe lane issue

Consumer platforms are also limited.

Example:

Intel i9-13900K

CPU PCIe:

CPU
 |
 +-- PCIe 5.0 x16
 |
 +-- PCIe 4.0 x4

So motherboard manufacturers split:

GPU1: x8
GPU2: x8

or:

GPU1: x16
GPU2: chipset x4

depending on design.

For AI workloads:

LLM inference:
- PCIe bandwidth matters less
- multiple GPUs can work
Training:
- gradient synchronization matters
- PCIe bandwidth matters a lot

Example multi-GPU setup

Cheap AI workstation

ATX motherboard

RTX3090
PCIe x8

RTX3090
PCIe x8

128GB RAM

850W~1200W PSU

Good.

mATX with risers

Possible:

mATX

GPU1
 |
PCIe riser

GPU2
 |
PCIe riser

but:

worse airflow
fewer lanes
more cable problems
harder debugging

It is common in mining, less ideal for AI.

Server/workstation boards solve this

For 4-8 GPUs:

EPYC motherboard

CPU PCIe lanes:

128 lanes

GPU1 x16
GPU2 x16
GPU3 x16
GPU4 x16
GPU5 x16
GPU6 x16
GPU7 x16
GPU8 x16

Examples:

AMD EPYC
Intel Xeon W
Xeon Scalable

This is why AI servers use boards like Gigabyte G292-Z20, Supermicro, ASUS ESC series.

For your RTX 4070 / possible RTX3090 multi-GPU experiments:

2 GPUs → ATX consumer board is reasonable
3-4 GPUs → workstation/server board is much better
8 GPUs → EPYC + PCIe risers/backplane

PCIe risers are mainly a mechanical solution. They do not create more PCIe lanes. The motherboard + CPU determine the actual bandwidth.

Back Donate