Dual GPU Installation Guide for AI

Dual GPU Installation Guide for AI | Generated by AI

Home 2026.07

Question: How to install two RTX 4070 or RTX 3090 GPUs together with an MSI motherboard? Is mATX too small? Which MSI boards are suitable?

Answer:

Yes, most mATX boards are a bad choice for dual large GPUs, especially RTX 3090. The issue is not only PCIe slots — it is physical spacing, airflow, power, and PCIe lane layout.

For AI workloads (PyTorch, vLLM, training), you want:

Two full-length PCIe x16 physical slots
At least 3-slot GPU spacing
Good VRM/power delivery
ATX/E-ATX case
Enough PSU capacity

1. Why mATX is problematic

Typical MSI mATX board:

CPU
 |
PCIe x16  ---- RTX 3090 (3 slots)
 |
PCIe x1
 |
PCIe x16 ---- second GPU

A RTX 3090 is usually:

Slot 1: GPU
Slot 2: cooler
Slot 3: cooler thickness

After installing the first card:

[ RTX 3090 ]
[ RTX 3090 cooler ]
[ blocked PCIe slots ]

The second GPU gets:

almost no airflow
temperature +15~30°C
possible thermal throttling

For RTX 4070 it is easier because most are 2-slot cards, but still tight.

2. MSI boards I would consider

Option A: MSI MAG B760 Tomahawk WiFi (ATX)

MSI MAG B760 Tomahawk WiFi

Good balance:

ATX size
2 × PCIe x16 physical slots
DDR5
supports 13th/14th gen Intel CPUs
good VRM

Layout:

CPU

PCIe x16
   |
   |  RTX 3090

3 slots gap

PCIe x16
   |
   |  RTX 3090

This is probably the minimum I would buy.

Option B: MSI PRO Z790-A MAX WiFi

Better for AI builds:

more PCIe flexibility
stronger VRM
better memory support

Good with:

i7-13700K
i7-14700K
i9-13900K

Option C: MSI MAG Z790 Tomahawk MAX WiFi

Higher-end choice.

Advantages:

stronger power design
more expansion
better if running two GPUs continuously

3. The real limitation: PCIe lanes

Important:

Most consumer Intel CPUs:

CPU PCIe lanes:

16 lanes GPU
4 lanes NVMe

Two GPUs become:

GPU1: x8
GPU2: x8

Example:

RTX 3090 #1
PCIe 4.0 x8

RTX 3090 #2
PCIe 4.0 x8

For LLM inference/training:

Usually OK.

You lose some bandwidth, but VRAM matters much more.

4. Power supply

Two RTX 3090:

GPU:

3090 ≈ 350W each

Two:

700W

CPU:

150-250W

System:

100W

Real requirement:

1000W minimum
1200W recommended

Good choices:

Corsair HX1200
Seasonic PRIME 1300W
MSI MPG A1000G/A1250G

5. Case matters more than motherboard

For dual 3090:

Avoid normal gaming cases.

You want:

8-slot GPU support
420mm GPU clearance
front mesh airflow

Examples:

Fractal Design Define 7 XL

Lian Li O11D XL

Server chassis is even better.

6. For your AI use case

Given you are thinking about:

local LLM
Qwen/DeepSeek style models
fine-tuning
vLLM

I would rank:

Best value

2 × RTX 3090 24GB
+
MSI B760 Tomahawk
+
128GB DDR5 RAM
+
1200W PSU

Total VRAM:

48GB

This is a very capable local AI box.

Cheaper path

2 × RTX 4070 12GB

VRAM:

24GB total

But for LLM:

48GB 3090 > 24GB 4070

because memory capacity dominates.

For your existing i9-13900KF + MSI B760M Mortar + RTX 4070, the first upgrade is actually motherboard + case + PSU. The B760M Mortar is a good gaming board, but it is not designed as a dual-GPU AI workstation.

Back Donate