Motherboard and CPU for a Multi-GPU Local LLM Rig (2026)

Bottom Line (August 2026)

• You do not need a Threadripper. HEDT platforms buy CPU lanes you only need at four-plus GPUs or for training. Two cards do not justify the price.
• You do need to read the slot table. Two physical x16-length slots does not mean x8/x8. On a B650 board the second slot is typically chipset-connected at x4.
• The phrase to search for on any board spec sheet is an explicit “x8/x8” mode on CPU-connected slots. X670E and X870E creator boards document it; mainstream B650 boards generally do not.
• Your second NVMe can cost you the second GPU slot. On the ASUS ProArt X870E-CREATOR, enabling the second CPU M.2 drops slot 2 from x8 to x4.
• The CPU barely matters for inference that fits in VRAM. It matters a lot the moment you offload to system RAM.
• Building one GPU? Ignore all of this. A B650 board is the correct, cheap answer and the money belongs in the GPU.

Start Here: Are You Actually Building Multi-GPU?

Most people reading this should build a single-GPU machine.

A second card is worth it for exactly one reason: capacity you cannot buy any other way. Two used RTX 3090s give you 48GB for roughly $2,000-2,600, which no single card matches at that price. That is a real argument and we make it in our 48GB VRAM guide.

What a second card does not give you is speed. Splitting a model across two GPUs does not make token generation twice as fast — in a layer split, the cards take turns. You buy VRAM, not throughput.

So if one card already holds your model, stop reading and put the money into a bigger card. If it does not, continue.

The Slot Table Is the Whole Post

Here is the verified specification wording for three AM5 boards, taken from the manufacturers’ own spec pages.

That is the whole difference, and it is not visible from a product photo. Both boards have more than one x16-length slot. Only one of them can give both slots to the CPU.

On the B650 board, the second x16-length slot hangs off the chipset and runs x4. Everything reaching that card crosses the chipset uplink, which is shared with USB, SATA and the chipset M.2 slots. It works. It is not what you buy a dual-GPU board for.

A note on how to read spec sheets: manufacturers list slot behaviour per CPU generation, and the answer changes. The same ASUS TUF B650 board gives Ryzen 8000-series chips a CPU slot that runs in x8/x4 mode, while Ryzen 9000 and 7000 get the full x16. Always find the row for the CPU you are actually buying.

The Trap Nobody Writes About: Your SSD Eats a GPU Slot

CPU PCIe lanes are a fixed budget, and a board’s marketing counts them optimistically.

From the ProArt X870E-CREATOR specification: the second CPU-connected M.2 slot shares bandwidth with the second x16 slot. Enable M.2_2, and slot 1 runs at x8 while slot 2 drops to x4. You wanted x8/x8; you have x8/x4 because you installed a second fast NVMe drive.

This is why a build can be correct on paper and disappointing in practice. The fix is boring and cheap: put your second and third drives in the chipset M.2 slots, which on that board are PCIe 4.0 x4 and do not touch the CPU’s graphics lanes. Model files are read once at load time — see our storage sizing notes — so a Gen4 chipset slot is genuinely fine for your model library.

Read the storage section’s footnotes on any board before you commit. That is where the lane conflicts live.

Does x4 Actually Hurt Inference?

This is where most guides quote a benchmark number, and the numbers in circulation contradict each other badly. We are not going to add to that. Here is the mechanism instead, which you can reason about without trusting anyone’s chart.

During token generation, a layer-split model passes a single activation tensor from one card to the next, once per layer boundary, per token. That is kilobytes. Even a PCIe 3.0 x4 link moves kilobytes in negligible time compared to the memory reads happening inside each GPU. This is the same argument that makes external GPUs less catastrophic for inference than for gaming.

During model loading and prefill, the picture changes. Loading tens of gigabytes of weights across a x4 link is measurably slower than across x8, and long prompts push far more data across the link than single-token decode does. If you run long-context agent sessions, prefill is a real part of your wall-clock time — our context window traps post covers why.

So: x4 is survivable for chat, x8 is what you want for agents and long prompts, and neither is the reason to buy a $700 board. Buy x8/x8 because it costs little on top of a board you were buying anyway, not because x4 will ruin you.

The CPU: Spend Less Than You Think

For inference that fits in VRAM, the CPU sets up the work and gets out of the way. It tokenises, samples, and schedules. The heavy arithmetic and the bandwidth-bound weight reads all happen on the GPU. A current mid-range chip is enough, and core count is close to irrelevant.

There is one important exception, and it is becoming more common: partial offload. When a model does not fit in VRAM, llama.cpp can keep some layers in system RAM, and then your CPU and your memory subsystem set the speed. This is how people run large MoE models on modest cards — see MoE vs dense on 24GB and running a 128GB-RAM model with no GPU. If that is your plan, prioritise memory channels and RAM capacity over cores.

If you want one CPU that covers both cases without a second thought, the Ryzen 9 9950X is the top of the AM5 mainstream stack and removes the question.

CPUAMD Ryzen 9 9950X ↗

What to Buy

If you are building one GPU

A B650 board is exactly right, and spending more is spending it in the wrong place. Both of these are solid AM5 boards with a full-speed CPU x16 slot on Ryzen 9000/7000.

B650ASUS TUF Gaming B650-PLUS WIFI ↗

B650MSI MAG B650 Tomahawk WIFI ↗

If you are building two GPUs

Buy an X670E or X870E board whose specification explicitly documents an x8/x8 mode, such as the ASUS ProArt X870E-CREATOR WIFI. We do not currently carry a link for a creator-class X870E board, so this is a plain-text recommendation rather than a placed one — verify the slot table yourself against the CPU you are buying.

Then check three physical things the spec sheet will not tell you:

1. Slot spacing. Most 24GB and 32GB cards are 2.5 to 3 slots thick. Two of them need real clearance, or the top card suffocates.
2. Case volume and airflow. Two 3090s put roughly 700W of heat into the room. Our quiet-cooling notes price what that costs to run.
3. Power. A dual-GPU build needs a 1200W-class ATX 3.1 unit. This is not the place to economise — see what PSU a local AI rig needs.

1200WMSI MAG A1200PLS, ATX 3.1 dual-GPU PSU ↗

CASELian Li Lancool III, high-airflow ↗

And the cards themselves:

24GBEVGA RTX 3090 24GB ↗

32GBGIGABYTE RTX 5090 WINDFORCE 32GB ↗

The Order to Spend In

1. VRAM. Everything else is downstream of what fits.
2. PSU. The one component where being wrong is destructive rather than slow.
3. Board, if and only if you are going dual. An x8/x8 board over a B650 is a small premium on a build that already costs thousands.
4. RAM. Enough to hold a model while it loads, and much more if you plan to offload. The DRAM shortage has made this line hurt.
5. CPU. Last. It is not your bottleneck.

That order is the actual advice. The Threadripper question people arrive with is a distraction from it.

See Also

• Best 48GB VRAM Setup for Local LLMs — the reason to go dual in the first place
• Dual RTX 3090 vs RTX 5090 — 48GB slow against 32GB fast
• What PSU Do You Need for a Local AI Rig? — transient spikes, and why gaming wattage advice is wrong
• Best Budget Local AI PC Under $1,000 — the single-GPU parts list
• Can I Run a Local LLM with 128GB RAM and No GPU? — the case where the CPU does matter
• Local LLM Electricity Cost Break-Even — what 700W of GPU costs to run
• Can You Put Four RTX PRO 6000 Max-Q Cards in One Workstation? — where 64 PCIe 5.0 lanes and eight slot widths are the requirement