Z790-P Multi-GPU PCIe Limitations

Z790-P Multi-GPU PCIe Limitations | Generated by AI

Home 2026.07

Question: How good is the PCIe layout of MSI PRO Z790-P for multi-GPU / AI workloads?

Answer:

The MSI PRO Z790-P has a consumer desktop PCIe layout, not a workstation/server multi-GPU layout. It is good for one powerful GPU + expansion cards, but not ideal for 2× RTX 3090 / AI training rigs.

The slots:

CPU PCIe lanes:
PCI_E1:
    PCIe 5.0 x16
    ↑ directly connected to Intel CPU

Z790 chipset lanes:
PCI_E3:
    PCIe 4.0 x4

PCI_E4:
    PCIe 3.0 x1

PCI_E2:
    PCIe 3.0 x1

Physical layout:

CPU
 |
 |  PCIe 5.0 x16
 |
[ GPU slot ]  <- RTX 4070 / 3090 here

(blank space)

[ PCIe x4 slot ] <- second GPU possible but only x4

[ PCIe x1 ]

[ PCIe x1 ]

For RTX 3090 + RTX 3090

Example:

GPU1

RTX 3090:

PCIe 4.0 x16

Actually runs:

PCIe 4.0 x16

Good.

GPU2

RTX 3090:

PCI_E3:
PCIe 4.0 x4

Bandwidth:

PCIe 4.0:
1 lane ≈ 2 GB/s each direction

x4:
≈ 8 GB/s

Compared with x16:

x16:
≈ 32 GB/s

So second GPU has only 25% bandwidth.

For inference:

GPU1 -> model layers
GPU2 -> model layers

PCIe x4 is usually OK.

For training:

GPU1 gradients
       |
       | PCIe
       |
GPU2 gradients

x4 becomes painful.

The bigger issue: Intel desktop CPU lanes

Your i9-13900KF:

CPU PCIe lanes:
16 lanes GPU
4 lanes NVMe

Only:

GPU:
x16

There is no:

x16/x16

support.

The motherboard cannot create lanes.

A Z790 board does not magically give more GPU bandwidth.

Compare better boards

For your AI use case (3090, 4070, ROCm experiments), look for:

MSI MPG Z790 Carbon WiFi

Better:

GPU1:
PCIe 5.0 x16

GPU2:
PCIe 5.0 x8

because it uses CPU bifurcation.

Both GPUs get CPU lanes:

x8 + x8

Much better.

ASUS ProArt Z790 Creator WiFi

More workstation-like:

PCIe 5.0 x16
PCIe 5.0 x8
Thunderbolt
10Gb Ethernet

Good for creators / AI.

Real multi-GPU platform

If your goal becomes:

2× RTX 3090
4× RTX 3090

then consumer Intel is the wrong platform.

Need:

AMD Threadripper
Intel Xeon W
EPYC

Example:

Threadripper:

CPU PCIe lanes:
64-128 lanes

Then:

GPU1 x16
GPU2 x16
GPU3 x16
GPU4 x16

For your current situation

You have:

i9-13900KF
RTX 4070
AI training interest
considering RTX 3090 second hand

I would rank options:

Option A — cheapest

Keep PRO Z790-P:

RTX 4070:
PCIe x16

RTX 3090:
PCIe x4

Good for:

LLM inference
QLoRA
vLLM experiments
running 40B+ models

Not optimal for:

distributed training

Option B — change motherboard

Buy Z790 with x8/x8:

RTX 3090
+
RTX 3090

both x8

Much cleaner.

Option C — workstation

Later:

Threadripper + 4 GPUs

when you really need multi-GPU training.

For your use case (training GPT-2 class models, experimenting with 24GB VRAM), the PRO Z790-P is acceptable. The second RTX 3090 at PCIe 4.0 x4 will hurt training throughput, but it will not stop you from running larger models. For inference, VRAM matters far more than PCIe bandwidth.

Back Donate