The short version: any ATX motherboard with two x16-length PCIe slots spaced at least three slots apart will work for dual-GPU LLM inference. The second slot running at x8 is fine — PCIe bandwidth is not the bottleneck for token generation. But physical slot spacing is critical, and that’s where most boards fail.
NVIDIA GeForce RTX 3090
24GB GDDR6XTwo used RTX 3090s give 48GB combined VRAM for ~$1,640. Make sure your motherboard can physically fit both cards.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
What actually matters for dual-GPU LLM
There are three things that matter for a dual-GPU LLM motherboard. Everything else — RGB, WiFi version, audio chipset — is irrelevant to inference performance.
1. Physical slot spacing
RTX 3090s and 4090s are triple-slot cards. Some aftermarket coolers are 3.5 slots. If your motherboard puts the second x16 slot directly below the first with only two slots of clearance, two large GPUs physically will not fit.
You need at least 3 slots of clearance between your two x16 PCIe slots. Many ATX boards space them at slots 1 and 4 (counting from top), which gives exactly 3 slots — tight but workable with most reference-design cards. Some boards space them at slots 1 and 5, which is ideal.
Check the motherboard manual’s physical layout diagram before buying. Photos of completed builds on forums are even better.
2. PCIe lane allocation
Consumer CPUs have limited PCIe lanes:
| Platform | CPU PCIe Lanes | Typical Dual-GPU Config |
|---|---|---|
| AMD AM5 (Ryzen 7000/9000) | 28 lanes (24 usable) | x16 / x8 or x8 / x8 |
| Intel LGA 1700 (13th/14th gen) | 20 lanes (16 usable) | x8 / x8 |
| Intel LGA 1851 (Arrow Lake) | 24 lanes (16 usable) | x8 / x8 |
| AMD TRX50 (Threadripper) | 48 lanes | x16 / x16 |
Does x8 vs x16 matter for LLM inference? Barely. Token generation is bound by memory bandwidth on the card, not by the link to it. The model weights sit in GPU VRAM and stay there; what crosses PCIe during inference is small, mostly activations handed between cards when the model is split. That is why an x8/x8 split is the normal dual-GPU configuration rather than a compromise — Intel lists 2x8+2x4 among the supported lane configurations for Arrow Lake, and AM5 boards bifurcate the same way.
Don’t overspend on a HEDT platform just for PCIe lanes unless you plan to run 3+ GPUs.
3. Power delivery to slots
Each PCIe x16 slot delivers up to 75W to the GPU. The rest comes from PSU cables. This is handled by the PSU, not the motherboard — but cheap boards with thin PCB traces near the slot can cause issues under sustained load. Stick with mid-range or higher boards from known manufacturers.
Check NVIDIA GeForce RTX 4090 on Amazon→Buy on Shopee SG→Best motherboards by scenario
Budget dual 3090 build (~$150-200 board)
Any mid-range ATX X670E or B650E motherboard with two x16-length slots will work. Popular, proven options:
- ASUS TUF Gaming X670E-PLUS — slots 1 and 4, x16/x4 split (x4 is fine for inference)
- MSI MAG X670E TOMAHAWK WIFI — good slot spacing, single M.2 between GPU slots
- Gigabyte X670E AORUS MASTER — slots 1 and 5, generous spacing
For Intel, the MSI MAG Z790 TOMAHAWK and ASUS TUF Z790-PLUS are reliable picks at similar prices. Both support x8/x8 bifurcation for dual GPUs.
Dual 4090 build (~$200-350 board)
The same boards work, but verify two things: slot spacing accommodates 4090 cooler size (some 4090s are 3.5 slots), and your case has adequate airflow for 900W of GPU heat. The multi-GPU setup guide covers case selection.
Serious multi-GPU (3-4 cards) build
Move to HEDT. AMD TRX50 (Threadripper 7000) motherboards like the ASUS Pro WS TRX50-SAGE offer 7 PCIe x16 slots and 48 CPU lanes. This is overkill for dual GPU but necessary if you’re scaling to 4x cards for 70B+ model inference.
Check NVIDIA GeForce RTX 3090 on Amazon→Buy on Shopee SG→Common mistakes to avoid
Buying a Micro-ATX board. Most mATX boards have only one x16 slot, or two slots so close together that two triple-slot GPUs can’t physically fit. Stick with full ATX or E-ATX.
Assuming both slots run at x16. On consumer platforms, the second slot almost always drops to x8 or x4 when both are populated. This is normal. Check your motherboard manual for the “dual GPU” lane allocation.
Ignoring M.2 slot conflicts. Some boards disable the second PCIe x16 slot when certain M.2 slots are populated, because they share PCIe lanes. Read the manual’s footnotes — the conflict table is usually buried on page 30.
Forgetting BIOS settings. Some boards default to “auto” lane allocation that gives the second slot x4 instead of x8. You may need to manually set PCIe bifurcation in BIOS. For the full dual 3090 setup walkthrough, I cover BIOS configuration step by step.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
Quick decision guide
| Your situation | Recommended board | Budget |
|---|---|---|
| Dual 3090, tight budget | B650E ATX (AM5) | ~$150 |
| Dual 3090/4090, mainstream | X670E ATX (AM5) | ~$200-250 |
| Dual 4090, max airflow | X670E or Z790 ATX | ~$250-350 |
| 3-4 GPUs, production server | TRX50 (Threadripper) | ~$800+ |
My pick for most dual-GPU LLM builders: a $200 X670E ATX board with a Ryzen 7 7700X. It’s inexpensive, has enough lanes, and the AM5 platform will be supported through 2027+. Pair it with the right PSU and you’re set.
NVIDIA GeForce RTX 4090
24GB GDDR6X48GB combined VRAM handles 70B models. Just make sure your motherboard has 3+ slots of spacing between PCIe x16 slots.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Frequently asked questions
Does the second PCIe slot running at x8 hurt LLM performance?
No. LLM inference is compute and VRAM bandwidth-bound, not PCIe bandwidth-bound. The difference between x16 and x8 is under 3% for token generation.
Can I use a Micro-ATX board for dual GPUs?
Not recommended. Most mATX boards lack the physical spacing for two triple-slot GPUs. Use a full ATX or E-ATX motherboard.
Do I need a Threadripper for dual GPU LLM?
No. Consumer AM5 or LGA 1700/1851 platforms handle dual GPUs fine for inference. Threadripper is only necessary for 3+ GPUs or if you need x16/x16 lanes.