AI

Best Motherboards for a Multi-GPU AI Build: AM5, TRX50, WRX90

This post contains affiliate links. If you buy through them, we may earn a small commission at no extra cost to you. Learn more.

Your second GPU shows up as x4 in nvidia-smi, the cards touch, and the top one cooks. That is what happens when you pick a motherboard for a multi-GPU AI build by the number of x16-sized slots instead of how they are wired and spaced. This guide picks the board by three numbers: how many lanes each card really gets from the CPU, how far apart the slots sit, and what the memory will cost you on that platform.

Original content from computingforgeeks.com - post 176554

We checked every board against the maker manual and the Amazon US listing on October 21, 2026, and the only performance numbers here come from our rented dual RTX 3090 test, an X99 host on vast.ai with Gen3 x16 per card. No board below went through our lab, so slot positions come from the manual drawings and lane wiring from the spec pages.

See the quick picks

Each board wins one kind of build. Start with the number of GPUs you plan to run:

BoardPlatformGPUs at full CPU widthPrice band
ASRock X870E Taichi LiteAM52 at x8/x8, room for 3-slot cards$290 to $320
ASUS ProArt X870E-Creator WiFiAM52 at x8/x8, best with 2-slot cards$500 to $550
MSI MEG X870E GODLIKEAM52 at x8/x8, room for 3-slot cards$1,050 to $1,150
ASUS Pro WS TRX50-SAGE WIFI ATRX503 at Gen5 x16 plus 1 at Gen4 x16$900 to $950
Gigabyte TRX50 AERO DTRX502 at Gen5 x16 plus 1 at Gen4 x16$520 to $560
ASUS Pro WS WRX90E-SAGE SEWRX904 at x16 (2-slot cards)$1,150 to $1,300

Prices moved a lot in 2026, and the Gigabyte board was only sold by a third-party seller on Amazon that day, so check the current price and seller before you order.

Know how we picked

A board made this list only if it passed four checks, and most popular gaming boards failed the first one:

  • Both GPU slots wired to the CPU, not the chipset
  • Slot positions read from the manual drawing
  • A buyable new listing on Amazon US
  • Maker spec page or manual for every number

We left out the MSI MPG X870E Carbon WiFi, because its second slot runs at x4, and the original TRX50-SAGE WIFI, which cannot run two 3-slot cards at Gen5 x16/x16. The ASRock TRX50 WS was out of stock, and the ASRock WRX90 WS EVO was listed only used, so neither could be linked.

Compare the boards

This is the table to keep open while you plan the case and the cards:

BoardForm factorSlot gap for card 2Case openings neededLANMemory
Taichi LiteEATX4 positions (one free slot)9 (two 3-slot)5GbE2-ch DDR5, ECC UDIMM
ProArt X870EATX3 positions (cards touch)7 (two 3-slot), 6 (two 2-slot)10GbE + 2.5GbE2-ch DDR5, ECC UDIMM
GODLIKEE-ATX4 positions8 (two 3-slot)10GbE + 5GbE2-ch DDR5, non-ECC
TRX50-SAGE WIFI ASSI CEBx16 slots at 1, 3, 5, 77 (two 3-slot), 8 (four 2-slot)10GbE + 2.5GbE4-ch RDIMM
TRX50 AERO DE-ATXZero gap at x16/x167 (two 3-slot), 8 (three 2-slot)10GbE + 2.5GbE4-ch RDIMM
WRX90E-SAGE SESSI EEB1/3/5/7 or 1/4/79 (three 3-slot), 8 (four 2-slot)Dual 10GbE + BMC8-ch RDIMM

1. ASRock X870E Taichi Lite: best value for two GPUs on AM5

ASRock X870E Taichi Lite motherboard
ASRock X870E Taichi Lite. About $290 to $320 on Amazon; check the current price.

The Taichi Lite is the board we picked for our dual 3090 build, and it is still the value pick. Both slots are CPU-fed PCIe 5.0, split x8/x8, and they sit four positions apart, so two 3-slot cards leave one free slot of air between them. No M.2 slot steals lanes from the second GPU slot.

The catches are networking and the case. It has 5GbE only, and the second card covers positions 7 to 9, so you need a case with nine expansion openings; that build’s parts list uses a Phanteks Enthoo Pro 2 Server Edition with 11. There is no third slot, so plan on two GPUs for the life of the board. It supports ECC UDIMMs.

2. ASUS ProArt X870E-Creator WiFi: best for 2-slot cards and 10GbE

ASUS ProArt X870E-Creator WiFi motherboard
ASUS ProArt X870E-Creator WiFi. About $500 to $550, sold by Amazon; check the current price.

Do you want 10GbE in a normal ATX case? The ProArt X870E-Creator gives you 10GbE plus 2.5GbE, dual USB4 and x8/x8 from the CPU. Its GPU slots sit three positions apart, so two 2-slot cards get one gap, but two 3-slot cards touch.

Watch the M.2 sharing too: putting a drive in M.2_2 drops the first slot to x8 and the second to x4. Our local AI workstation build used this board with a riser to move the second 3-slot card away. ASUS lists 256GB on the web page and 192GB in one manual revision, so check the QVL before buying four 64GB sticks.

3. MSI MEG X870E GODLIKE: best premium AM5 board

MSI MEG X870E GODLIKE motherboard
MSI MEG X870E GODLIKE. About $1,050 to $1,150, sold by Amazon; check the current price.

The GODLIKE combines what the first two picks split: four-position slot spacing for 3-slot cards and 10GbE plus 5GbE. The second card covers positions 6 to 8, so an 8-opening case is enough, and it sits over the chipset x4 slot.

It supports non-ECC memory only, and it costs about as much as a TRX50 board that runs three cards at full x16. Buy it if you want to stay on AM5 for the cheaper CPU and memory, but need fast networking and proper card spacing in one board.

4. ASUS Pro WS TRX50-SAGE WIFI A: best Threadripper board for 3 to 4 GPUs

ASUS Pro WS TRX50-SAGE WIFI A motherboard
ASUS Pro WS TRX50-SAGE WIFI A. About $900 to $950, sold by Amazon; check the current price.

The TRX50-SAGE WIFI A has three PCIe 5.0 x16 slots and one PCIe 4.0 x16, all wired to the CPU on every Threadripper. The x16 slots sit at positions 1, 3, 5 and 7, so two 3-slot cards fit at full Gen5 x16/x16 with a gap, and four 2-slot cards fit at x16.

It adds 6-pin and 8-pin connectors for slot power, and ASUS says to connect the 8-pin, or both, never the 6-pin alone. You get 10GbE plus 2.5GbE and a BMC header, but the IPMI card is sold separately. Make sure the listing says WIFI A: the original 2023 board puts its second x16 slot two positions down, so two 3-slot cards cannot both run at Gen5 x16.

5. Gigabyte TRX50 AERO D: cheapest way onto Threadripper lanes

Gigabyte TRX50 AERO D motherboard
Gigabyte TRX50 AERO D. About $520 to $560 from a third-party seller; check the current price.

The TRX50 AERO D is the cheapest board here that gets you Threadripper lanes. It has two PCIe 5.0 x16 slots and one PCIe 4.0 x16, all from the CPU, plus 10GbE and a standard E-ATX depth.

Two 3-slot cards fit at x16/x16 only with zero gap, so plan on 2-slot or blower cards for three GPUs. There is no extra slot power connector, and on October 21 Amazon had it only from a third-party seller with two units left, so confirm the seller before you buy.

6. ASUS Pro WS WRX90E-SAGE SE: best for four Gen5 x16 GPUs and 8-channel memory

ASUS Pro WS WRX90E-SAGE SE motherboard
ASUS Pro WS WRX90E-SAGE SE. About $1,150 to $1,300, sold by Amazon; check the current price.

The WRX90E-SAGE SE is the board for four GPUs all at Gen5 x16. ASUS lists a 4-way layout in slots 1, 3, 5 and 7 and a 3-way layout in slots 1, 4 and 7, all at x16, which fits four 2-slot cards or three 3-slot cards. It adds 8-channel RDIMM memory, an AST2600 BMC with its own port, dual 10GbE and two 8-pin slot power connectors.

It takes Threadripper PRO WX chips only, not the non-PRO parts that fit TRX50. It is an SSI-EEB board, so the case must take EEB, and a 3-way build with 3-slot cards needs nine openings. Eight RDIMMs to fill the channels are the real cost, which Step 4 covers.

Step 1: Count the CPU lanes before you pick a platform

Lanes from the CPU decide how many cards run at full width, and our PCIe lanes guide covers how they are split. These are the usable lane counts AMD lists on its Threadripper platform page, its 2023 Threadripper 7000 launch release and the Ryzen 9000 spec:

PlatformUsable CPU lanesGPUs at full CPU widthMemory
AM5 (Ryzen 9000)24, PCIe 5.02 at x8/x82-ch UDIMM, 256GB
TRX50, Threadripper 700088, up to 48 PCIe 5.02 to 3 at Gen5 x164-ch RDIMM, 1TB
TRX50, Threadripper 900088, up to 80 PCIe 5.03 to 4 at x164-ch RDIMM, 1TB
WRX90, Threadripper PRO144, 128 PCIe 5.04 at x168-ch RDIMM, 2TB

A third card on AM5 hangs off the chipset at x4 or x2, behind the single chipset uplink, so treat AM5 as a two-GPU platform. Pair each board with one of these CPUs, priced on Amazon on October 21:

CPUPlatformPrice bandNote
Ryzen 9 9950XAM5about $540Sold by Amazon
Threadripper 7960XTRX50about $1,100Up to 48 Gen5 lanes, in stock
Threadripper 9960XTRX50about $1,450Up to 80 Gen5 lanes, later ship date
Threadripper PRO 9955WXWRX90about $1,550Cheapest PRO listing we found, third-party seller
Threadripper PRO 9975WXWRX90about $3,900Sold by Amazon, limited stock

Step 2: Check slot spacing against your cards

A 3-slot card in one slot needs the next GPU slot at least four positions lower to leave one free slot for air. At three positions the cards touch, and the top card breathes the exhaust of the one below it.

  • Taichi Lite, GODLIKE: 4 positions, one free slot
  • ProArt X870E: 3 positions, cards touch
  • TRX50-SAGE WIFI A: x16 slots at 1, 3, 5 and 7
  • TRX50 AERO D: zero gap at x16/x16

Then count case openings. A 3-slot card in position 7 covers positions 7 to 9, so it needs a case with nine expansion openings; a card in position 6 needs eight. A low card also sits over the bottom-edge front-panel, USB and audio headers, so plug those in before the card goes in. A riser is the escape hatch when spacing fails, but buy one rated for the slot’s PCIe generation and check the link width after you install it.

Step 3: Match the split mode to your lanes

Lanes matter most when the cards talk to each other on every layer. In our dual 3090 test, each card sat on PCIe Gen3 x16 with no NVLink, which is about the same bandwidth a Gen4 card gets in an AM5 x8 slot:

Llama 3.3 70B Q4_K_M (llama.cpp)Token generationPrompt processing
Layer split19.48 tok/s577 tok/s
Tensor split31.93 tok/s763 tok/s

Tensor split ran 1.64 times faster, so two cards at x8 are worth it on AM5. Layer split barely touches the bus, which is why it survives narrow links, while tensor parallel syncs every layer. We did not measure x8 against x16 ourselves, and we found no public test of it for tensor parallel, so we cannot give you a percentage.

With more cards the picture changes. The vLLM docs prefer pipeline parallelism when GPUs have no NVLink, and a public 4x RTX 3090 test reached 1,437 tok/s with two TP=2 copies against about 600 tok/s for one TP=4 instance at 128 concurrent requests. Our reading: on a four-GPU box, give every card a CPU-direct x16 slot, because the cards spend more time talking to each other.

Step 4: Budget memory and slot power

Memory now decides the platform as much as lanes do. On October 21, a single 32GB DDR5-6400 RDIMM cost about $1,400 from a third-party seller on Amazon, so four come to about $5,600; a Kingston 4x32GB DDR5-5600 kit, also third party, was about $4,200. Eight modules for WRX90 run about $8,400 to $11,200 at those prices. Our DRAM shortage explainer covers why.

On Threadripper, the memory bill now rivals the board and the CPU together, which pushes most two-GPU builds back to AM5 and its cheaper UDIMMs. Our ECC RAM guide explains UDIMM against RDIMM if you want ECC on AM5.

For power, a PCIe x16 slot is generally rated for up to 75 W, which is why the Threadripper boards add connectors: the TRX50-SAGE WIFI A has a 6-pin and an 8-pin, and the WRX90E-SAGE SE has two 8-pins. The AERO D, ProArt and Taichi Lite have none. Size the PSU from the cards, as our dual 3090 build does.

Do not trust the link speed at idle. GPUs drop to a low PCIe generation when idle, and one Level1Techs user saw a Taichi Lite with two RTX 5090s report Gen1 x8 at idle and Gen5 x8/x8 only under a tensor-parallel load. Run your model first, then check each card while it works:

nvidia-smi --query-gpu=index,pcie.link.gen.current,pcie.link.width.current --format=csv

Each card should report the generation and width its slot supports, such as 5 and 8 for a Gen5 card on AM5. To see the same thing from the kernel side, find each card’s PCI address first; nvidia-smi prints it as 00000000:01:00.0, and lspci needs only the 01:00.0 part:

nvidia-smi --query-gpu=index,pci.bus_id --format=csv

Then read the link status for that address:

sudo lspci -vv -s 01:00.0 | grep LnkSta

A width lower than the slot supports usually means an M.2 drive is sharing lanes, a riser is not seated, or the card is in a chipset slot. We list these commands from the NVIDIA and pciutils docs; the output depends on your cards.

What breaks first in a multi-GPU build

Most multi-GPU builds that go wrong fail on layout, not on the board itself. Every item below traces back to a number in this guide, so check each one against your parts list before you order:

  • Cards that touch and throttle the top GPU
  • An M.2 drive that drops the second slot to x4
  • A case one opening short for the bottom card
  • Four RDIMMs that cost more than the board
  • A third GPU on a chipset x4 slot

If you are still choosing the cards, our RTX 3090 vs 5090 comparison and the VRAM guide tell you how many you need before you pick the board.

Keep reading

Ollama Models Cheat Sheet: Best Models by VRAM, Tested (PDF) AI Ollama Models Cheat Sheet: Best Models by VRAM, Tested (PDF) Open Source LLM Comparison Table (2026) AI Open Source LLM Comparison Table (2026) Claude Code Cheat Sheet: Commands, Shortcuts and Flags (PDF) AI Claude Code Cheat Sheet: Commands, Shortcuts and Flags (PDF) OpenCode CLI Cheat Sheet: Commands, Sessions and Fixes (PDF) AI OpenCode CLI Cheat Sheet: Commands, Sessions and Fixes (PDF) Kimi K3.1 Pro for Coding: What Changed vs K3 and API Setup AI Kimi K3.1 Pro for Coding: What Changed vs K3 and API Setup Ollama Commands Cheat Sheet: CLI, API and Env Vars (PDF) AI Ollama Commands Cheat Sheet: CLI, API and Env Vars (PDF)

Leave a Comment

Press ESC to close