External GPU docks for local LLMs: OCuLink vs Thunderbolt/USB4

Your mini PC can keep its tidy little case and still put a desktop NVIDIA card to work. Choose MINISFORUM DEG1 for a native OCuLink host and a PSU you supply, AOOSTAR AG02 for OCuLink or USB4 with power included, or Sonnet Breakaway Box 750ex for an enclosed Thunderbolt setup. Then choose the card by VRAM: the dock connects the GPU; it doesn’t make the model smaller.

Specifications and setup guidance come from the makers, standards bodies and Ollama documentation, checked October 25, 2026.

Quick picks: start with the port, not the box

• Native OCuLink: MINISFORUM DEG1. An open dock with a direct PCIe 4.0 x4 connection and room for your own ATX/SFX supply. A good fit when your mini PC already has the right port.
• Two connection routes: AOOSTAR AG02. OCuLink and USB4, a built-in 800 W supply and an open design for long cards. Use the connection your host actually supports.
• Enclosed Thunderbolt: Sonnet Breakaway Box 750ex. A supplied 750 W PSU, covered card, USB ports and Ethernet. Sonnet’s current Windows specification calls for Windows 11 25H2 or later.
• 16 GB card: ASUS Dual RTX 5060 Ti OC. More room for the model and its context than the 12 GB MSI RTX 3060 Ventus 2X, which remains a smaller-model option. Our GPU buying guide explains the model-size choices.

If you’re choosing the whole computer too, read GPU desktop vs unified-memory mini PC first. An eGPU is especially appealing when you already like your host and want to add or reuse a card.

Compare the docks

Dock	Connection, power and fit
MINISFORUM DEG1: choose a direct OCuLink connection when your host has the port and you supply the PSU.	OCuLink, PCIe 4.0 x4. Open frame; supply your own ATX/SFX PSU and its cables. No USB4 route. Shut down before changing the OCuLink connection.
AOOSTAR AG02: choose two connection routes and an included supply when you want an open-card setup.	OCuLink + USB4. Built-in 800 W PSU; current maker summary specifies GPUs up to 600 W. Open card layout; up to 100 W USB4 charging. Maker recommends Windows 11.
Sonnet Breakaway Box 750ex: choose an enclosed Thunderbolt dock with USB ports and Ethernet for your desk.	Thunderbolt 3, with a PCIe 3.0 x4 electrical slot. Built-in 750 W PSU, supporting 375 W continuous GPU power plus 100 W peak. Two 6+2-pin GPU leads; card limit 312 × 160 × 55 mm. Four USB-A ports, Ethernet and up to 85 W laptop charging.

DEG1: a direct link and your own power supply

The DEG1 maker page specifies OCuLink PCIe 4.0 x4 and ATX/SFX support. The installation manual makes the shopping list explicit: the PSU, AC lead, 24-pin cable and GPU power cable are separate.

Use the modular cables supplied for that exact PSU, connect the dock and card as shown, and give the GPU clear air around its fans. Our PSU guide covers power and cable planning; here the mini PC keeps its own supply while the dock powers the card.

OCuLink is a power-off connection. DEG1’s manual forbids hot-swapping. Its follow-start behaviour is intended for compatible MINISFORUM hosts; for other hosts, the manual gives a dock-first, host-second startup sequence. Keep that sequence with your setup notes.

AG02: one dock for two kinds of host

The current AG02 specification gives you OCuLink and USB4 with the supply already installed. That suits a desk where you might change hosts without rebuilding the power side.

Use its 600 W GPU limit from the current specification summary, rather than treating the PSU’s 800 W label as a card allowance. Match the card’s power plugs to the supplied leads. OCuLink still requires shutdown before cable changes; the maker says so explicitly. Its compatibility notes also exclude Lenovo Legion Go, ROG Xbox Ally X and RX 9070 XT, so those combinations belong off this shortlist.

Sonnet 750ex: put the card in a proper enclosure

Sonnet’s current technical specifications separate PSU wattage, card power and physical clearance. That is the useful buying information: a card must fit all three dimensions, use the supported power connections and stay within the GPU allowance.

The ASUS card measures 229 × 120 × 50 mm, and the MSI measures 235 × 124 × 42 mm, both inside that dimensional envelope. Their makers specify one 8-pin input each. This is a physical-fit check; select the host OS and GPU driver from the enclosure and Ollama support lists too.

For the NVIDIA setup here, use the Windows route Sonnet documents. Apple-silicon Macs don’t support eGPUs; a Thunderbolt port on a Mac mini doesn’t change that. Sonnet’s installation guide covers card installation, auxiliary power and the supplied Thunderbolt cable.

How much bandwidth does the cable leave you?

OCuLink carries PCIe directly. Thunderbolt/USB4 carries PCIe through a tunnel alongside the link’s other traffic. Keep those layers separate when reading the big numbers on the box.

PCI-SIG’s generation table specifies 16 GT/s per lane for PCIe 4.0 and 8 GT/s for PCIe 3.0, both using 128b/130b coding. The calculations below use decimal GB/s, in one direction.

Path or requirement	Computed figure and what it means
PCIe 4.0 x4, as on DEG1’s OCuLink path	7.877 GB/s after line coding, before packet overhead.
PCIe 3.0 x4, Sonnet’s bridge-side slot	3.938 GB/s after line coding, before Thunderbolt overhead.
40 Gb/s Thunderbolt/USB4 transport	5.000 GB/s by bit-to-byte conversion; shared transport, not dedicated GPU payload.
Thunderbolt 3 / 4 minimum PCIe data requirements	2.000 / 4.000 GB/s, converted from Intel’s 16 / 32 Gb/s requirements. These are host minimums, not maximum copy speeds.

The PCIe calculation is GT/s × four lanes × 128/130 ÷ eight bits per byte. Intel confirms Thunderbolt 3 and 4 share 40 Gb/s transport while their minimum PCIe requirements differ. The USB-IF overview explains protocol tunnelling. Actual transfers also depend on the host, bridge, cable and competing traffic.

Loading the model is different from generating tokens

When Ollama loads a model onto the external GPU, its weights have to reach GPU memory across that connection. Change models often and you repeat that transfer. A faster direct link helps that part of the job.

Once the model is resident, the GPU reads its weights from its own GDDR memory while doing the inference work. It doesn’t ship the whole model across the dock cable for every new token. NVIDIA’s CUDA guidance explains why keeping data on the device avoids repeated host/device transfers; Ollama likewise favours fitting a model on one GPU to reduce PCI bus traffic.

That is why the cable matters much less for a fully resident model than for loading or an offload-heavy workload. CPU/GPU splits and multi-GPU communication change the traffic pattern. Link bandwidth is not a tokens-per-second multiplier.

VRAM first: your system RAM stays separate

A discrete GPU’s VRAM and the mini PC’s RAM remain separate physical memories. Ollama can split work between CPU and GPU, but adding a dock doesn’t turn system RAM into more GDDR.

Choose the ASUS Dual RTX 5060 Ti 16 GB when your model and context need the extra room. The ASUS specifications give its dimensions and single 8-pin input. The MSI RTX 3060 Ventus 2X 12G OC suits a smaller resident-model target; its maker sheet specifies 12 GB, 170 W and one 8-pin input.

Leave room for context and runtime allocations, using our Ollama context/VRAM guide. Both GPU families appear in Ollama’s NVIDIA support documentation. The dock doesn’t replace the correct driver.

Will your mini PC work? Run this checklist

1. Identify the real port. DEG1 needs OCuLink. A USB-C port that only provides USB data, charging or video won’t run an eGPU. For USB4, establish PCIe-tunnelling support and the dock’s host compatibility; for Sonnet, follow its Thunderbolt host requirements.
2. If using an M.2 adapter, inspect the slot and cable route. You need the PCIe/NVMe M-key route and the correct adapter direction. ADT-Link documents M.2-to-OCuLink adapters separately from the reverse direction. This occupies a storage slot and needs a safe way out of the case; a native OCuLink port is tidier.
3. Read the host firmware settings. The Linux kernel’s Thunderbolt/USB4 guide notes that some BIOS settings disable PCIe tunnelling and that devices may need authorization. Approve the intended dock through your OS’s normal flow, rather than bypassing security for every device.
4. Plan card size, plugs and power together. Include the dock’s card allowance, PSU leads and airflow. Don’t buy a long, thick GPU and hope the enclosure stretches.
5. Choose the host OS and driver before buying the dock. Keep the exact mini-PC specification handy; our mini-PC guide helps compare ports and expansion.

Set it up, then prove the model is on the GPU

With everything powered down, install the card and its required power leads. Follow the dock’s assembly and power-on sequence, using the supplied or maker-specified OCuLink/Thunderbolt/USB4 cable. Enable the required host PCIe path and authorize the dock if prompted. Install the supported NVIDIA driver, then confirm the card appears in nvidia-smi before troubleshooting Ollama.

Load a model you intend to use, then inspect residency:

ollama run MODEL_NAME

After the model responds, enter /bye to leave the chat, then run:

ollama ps

Replace MODEL_NAME with your installed model’s name. In PROCESSOR, 100% GPU means the model is entirely GPU-resident, not that GPU utilization is constantly full. A CPU/GPU split means some placement is in system memory. Reduce the model/context requirement or choose more VRAM if full residency is your goal.

Keep a favourite model resident while you use it, keep OCuLink cable changes to shutdowns, and pick the dock that fits your host. DEG1 for direct OCuLink, AG02 for two routes with power included, Sonnet for an enclosed Thunderbolt desk. Let VRAM choose the card.