AI-narrated version of this post using a synthetic voice. Great for accessibility or listening while busy.
A modern mini PC will happily run a 7B or 8B model under Ollama. Push past that and the integrated GPU runs out of memory bandwidth, the shared system RAM fills up, and token generation slows to the point where you stop using the thing. The obvious fix is a desktop graphics card, but mini PCs have no PCIe slot to put one in.
That is the gap OCuLink fills. It is a cable standard that carries raw PCIe lanes out of the chassis to an external dock, where a full-size card sits with its own power supply. For anyone running llama.cpp, Ollama or an image generation stack on a small-form-factor box, it is the closest you get to a real desktop GPU without buying a desktop.
This guide covers how OCuLink differs from Thunderbolt and USB4 enclosures, which mini PCs actually expose the port, how much VRAM you need for the model sizes you care about, and the practical annoyances nobody mentions until you have already bought the parts.
OCuLink is PCIe on a cable, nothing more
OCuLink (the connector family is formally SFF-8611) is a direct electrical extension of the host’s PCIe bus. There is no tunnelling protocol, no encapsulation, no controller chip translating one bus into another. The lanes that would have gone to an internal slot go out through a connector instead.
On mini PCs the port is almost always wired as four lanes. At PCIe 4.0 speeds that gives you roughly 64 Gbps of raw signalling, and because there is no protocol overhead sitting on top, most of it reaches the card. The GPU appears to the operating system as an ordinary PCIe device.
That simplicity is both the strength and the weakness. No controller means no negotiation, no power delivery, and no graceful handling of a cable being pulled out mid-session.
How it compares to Thunderbolt and USB4 enclosures
Thunderbolt 3, Thunderbolt 4 and USB4 all tunnel PCIe over a 40 Gbps link that is also carrying DisplayPort and USB traffic. The usable PCIe throughput after overhead lands well below the headline figure, which is why Thunderbolt eGPUs have a reputation for leaving gaming performance on the table.
For LLM inference the picture is less dramatic than you might expect. Once a model’s weights are resident in VRAM, generation speed is governed by the card’s own memory bandwidth, not the host link. Where the link matters is model loading time, swapping between models, and any scenario where layers spill back to system RAM.
So OCuLink wins clearly on load times and on partial-offload workloads, and wins modestly on everything else. Thunderbolt wins on convenience, and our comparison of eGPU enclosures goes deeper on the enclosure side if you are leaning that way.
| Connection | PCIe path | Typical lanes | Hot-plug | Powers the GPU? |
|---|---|---|---|---|
| OCuLink (SFF-8611) | Direct, no tunnelling | x4 | Not supported in practice | No, separate PSU required |
| Thunderbolt 3 / 4 | Tunnelled over 40 Gbps link | x4 equivalent, shared | Yes | Enclosure PSU, plus host charging |
| USB4 | Tunnelled, PCIe optional per host | Varies by implementation | Yes | Enclosure PSU, plus host charging |
| M.2 to PCIe adapter | Direct, via internal slot | x4 | No | No, separate PSU required |
| Internal desktop slot | Direct | x8 or x16 | No | Yes, from system PSU |
Which mini PCs actually expose an OCuLink port
The port is still a minority feature, concentrated in AMD-based enthusiast models. Check the spec sheet rather than assuming, because two machines in the same product family often differ.
The MINISFORUM AI X1 Pro-370 pairs a Ryzen AI 9 HX 370 with Radeon 890M graphics, 32GB of DDR5 and a 1TB PCIe 4.0 SSD, and carries both dual USB4 ports and an OCuLink port. Having both is useful: USB4 for docks and displays, OCuLink reserved for the GPU. Dual 2.5G LAN and Wi-Fi 7 round it out for anyone serving models to other machines on the network.
- Powerful AI Processor: Experience next-generation AI technology, greatly improve productivity, and bring unprecedente…
- Intelligent AI Assistant: Mini PC AI X1 Pro has a built-in new Copilot AI function and supports Recall function – jus…
- Extreme audio experience and efficient noise reduction: Equipped with dual noise reduction DMIC and built-in speakers…
The MINISFORUM UM890 Pro takes a similar approach around the Ryzen 9 8945HS. It lists one OCuLink port alongside two USB4 ports, four USB 3.2 Gen2 Type-A ports, dual 2.5G Ethernet and two M.2 2280 PCIe 4.0 slots. Memory expands to 96GB, which matters when you are running larger models partly on the CPU while the GPU handles what fits.
- 【Powerful Ryzen 9 8945HS】MINISFORUM UM890 Pro is equipped with a new-generation AMD AI Ryzen 9 8945HS processor, with…
- 【Large Memory & Storage Support】UM890 Pro is equipped with DDR5 32GB RAM,a SODIMM slot, frequency of up to 5600 MHz…
- 【Quad-Screen Display】Mini Pc Ryzen 9 8945HS UM890 Pro is equipped with HDMI 2.1, DP 1.4 , and 2 x USB4, supporting ma…
Not every capable mini PC has the port. The GMKtec K8 Plus, built on the Ryzen 7 8845HS with Radeon 780M graphics, 32GB of DDR5 and dual Intel i226V 2.5G NICs, connects external devices over USB4 instead. It is a solid host for a Thunderbolt-class enclosure or for iGPU-only inference, and if you want a broader view of what people are buying, our roundup of best-selling mini PCs covers the wider field.
- 【Advanced Gaming PC】Powered by the AMD Ryzen 7 8845HS processor with 8 C/ 16T, up to 5.1GHz, and featuring a 16MB L3…
- 【32GB DDR5 + 512GB SSD】Equipped with 32GB memory capacity,DDR5 speed up to 5600MHz, offering faster data processing a…
- WHY CHOOSE DDR5 5600MHz DUAL CHANNEL (2×16GB): With a 5600MHz clock—a 17% frequency uplift over 4800MHz—this kit deli…
VRAM tiers and what model sizes they hold
VRAM capacity decides which models you can run at all. Bandwidth decides how fast they generate. Capacity is the harder constraint, so start there.
The practical tiers
- 8GB: comfortable for 7B and 8B models at 4-bit quantisation, with room for a modest context window.
- 12GB: the same models at longer context, or 13B and 14B models quantised down.
- 16GB: 14B comfortably, 20B-class models with tighter quantisation.
- 24GB and above: 32B-class models, and the point where multi-user serving stops being painful.
Remember that the KV cache grows with context length and sits in VRAM alongside the weights. A model that loads fine at 4K context can run out of memory at 32K.
The ZOTAC GeForce RTX 3060 Twin Edge OC sits in that 12GB tier with a 192-bit GDDR6 bus running at 15 Gbps and a PCIe 4.0 interface. Ampere brings CUDA compatibility across essentially every local inference stack, which removes a whole category of driver headaches. If you are weighing bigger cards, we compared the options in our guide to high-VRAM GPUs for serious local AI.
- NVIDIA Ampere architecture, 2nd Gen Ray Tracing Cores, 3rd Gen Tensor Cores
- 12GB 192-bit GDDR6, 15 Gbps, PCIE 4.0; Boost Clock 1807 MHz
- IceStorm 2.0 Cooling, Active Fan Control, Freeze Fan Stop, Metal Backplate
Power, cable length and the no-hot-plug rule
OCuLink carries no power. The dock needs its own supply, usually an ATX unit or a dedicated brick, sized for the card’s rated draw plus transient spikes. Undersizing here produces crashes under load that look like driver faults.
Cable length is the second constraint. Passive OCuLink cables are specified for short runs, and signal integrity degrades quickly beyond that. Plan on the dock sitting next to the mini PC, not across the room.
Then the big one: no hot-plug. Power the dock and the GPU first, then boot the host. Pulling the cable on a running system will hang it. You may also need to enable Above 4G Decoding in the BIOS before the card is detected properly.
If adding a dedicated circuit or any in-wall wiring is part of your plan, follow local electrical code and use a licensed electrician where your jurisdiction requires one.
When a different approach fits better
If you move your machine around, Thunderbolt’s hot-plug support is worth more than OCuLink’s extra bandwidth. Docking and undocking without a reboot changes how you use the system day to day.
If you only ever run 7B or 8B models, an NPU-equipped mini PC on its own may be enough, and our look at Copilot+ PCs and NPU mini PCs explains where those chips help. And if your workload is bursty rather than constant, the maths in cloud GPU compute versus a home card may point somewhere else entirely.
Builders who want more than four lanes, or two cards, should skip all of this and buy a tower.
What to build
For a permanently sited local inference box, pair an OCuLink-equipped mini PC with a dock and a 12GB or larger card. The MINISFORUM UM890 Pro is the straightforward pick thanks to its OCuLink port, 96GB memory ceiling and dual M.2 slots. Choose the AI X1 Pro-370 if you want the newer HX 370 platform and Wi-Fi 7 alongside the same expansion path.
The RTX 3060 Twin Edge OC is a sensible first card for 7B to 14B models on CUDA. Budget for a properly sized PSU, accept the reboot ritual, and set the software side up following our walkthrough on self-hosting an AI chatbot with Ollama and Open WebUI.
How these picks were chosen: HomeNode compares products using manufacturer specifications, current Amazon listing details and published owner feedback. Products are not tested in-house, and listings change often, so confirm current specifications on the product page before buying.
Affiliate disclosure: As an Amazon Associate, HomeNode earns from qualifying purchases at no additional cost to you.
Related Auburn AI Products
Building a homelab or self-hosting content site? Auburn AI has practical kits: