An OCuLink dock lets a mini PC drive a full desktop graphics card for about $110. For local LLMs that fit in the card’s VRAM, the four-lane link costs you very little speed. The setups that suffer are the ones that spill model layers into system RAM, where prompt processing dropped by half in a Thunderbolt 3 test. That one rule decides which dock and which mini PC are worth your money.
This guide picks the best OCuLink eGPU docks and mini PCs with an OCuLink port for running Ollama or llama.cpp on an external GPU. It covers what the PCIe 4.0 x4 link really costs according to published benchmarks, which mini PCs give up an NVMe slot for OCuLink, and what to check before you order. Specs come from each manufacturer’s page, filled in from independent reviews where the maker is silent, and prices and stock were checked on Amazon on October 19, 2026.
Pick an OCuLink setup at a glance
You need two parts: a mini PC with an OCuLink port and a dock that holds the GPU and a desktop power supply. None of the docks here includes a PSU.
| Award | Pick | Why it wins | Price band |
|---|---|---|---|
| Best dock for most builds | Minisforum DEG1 | Open frame, any card length, PCIe 4.0 x4 | about $110 |
| Best dock for laptops too | Minisforum DEG2 | OCuLink plus Thunderbolt 5, M.2 slot, 2.5GbE (over TB5) | about $250 |
| Best dock that powers the mini PC | AOOSTAR EG02 | OCuLink plus Thunderbolt 5 and a 150 W DC line out | about $220 |
| Best value mini PC | GMKtec K12 | Dedicated OCuLink port and three M.2 slots | about $880 |
| Most proven mini PC | Minisforum UM890 Pro | Ryzen 9 8945HS, 400+ Amazon reviews | about $990 |
| Most RAM for CPU offload | Minisforum AI X1 Pro-470 | Takes 128GB, the only pick above 96GB | about $1,300 |
Prices are what the maker’s own Amazon store charged on October 19, and they move often, so check the live price before you order.
Know what the x4 link costs
The OCuLink 4i ports on these products carry four PCIe 4.0 lanes, about 7.9 GB/s each way, which is a quarter of a desktop x16 slot. In practice ChargerLAB measured 6.0 to 6.7 GB/s through a DEG1. Whether that matters depends on where the model’s weights live:
| Scenario | Narrow link | Full link | Result |
|---|---|---|---|
| Qwen3 30B Q4, fully in VRAM, RTX 4090 (generation) | PCIe 3.0 x1: 41.05 t/s | Desktop slot: 42.51 t/s | 3.4% slower |
| gpt-oss-120b, 12 of 36 layers on GPU (prompt processing) | TB3, PCIe 3.0 x4: 140 t/s | PCIe 3.0 x16: 304 t/s | 54% slower |
| Same model, token generation | 2.27 t/s | 2.33 t/s | 3% slower |
We took the first row from Jeff Geerling’s tests, which ran an RTX 4090 from a Raspberry Pi on a single PCIe 3.0 lane, an eighth of what OCuLink gives you, and still lost under 5% on a 30B model, because once the weights sit in VRAM, generation runs on the card. Multi-GPU setups behave the same way when the model is split by layers: an egpu.io test of two RTX 3090 docks ran a 70B model at 17.96 t/s with a DEG1 on OCuLink plus a USB4 dock, and 17.83 t/s with two daisy-chained Thunderbolt 3 docks. That compares two narrow links rather than x4 against x16, but it shows layer split does not lean on the link.
Partial offload is different. When some layers stay in system RAM, llama.cpp copies those weights to the GPU for every prompt batch of 32 tokens or more, so the link becomes the bottleneck. In the gpt-oss-120b test, a four-lane Thunderbolt connection halved prompt processing against an x16 slot on the same machine. That test used Thunderbolt 3, an effective 2.7 GB/s or so, and two different 3090 cards, so native OCuLink at 6 GB/s or more should lose less. The direction is still clear: buy enough VRAM for the models you run, and treat the x4 link as fine for that case.
Model loading is the other cost. A 20 GB model file needs at least 2.5 seconds of link time at x4 against about 0.6 seconds at x16, which is arithmetic, not a measurement, and on most mini PCs the NVMe drive is slower than either. Our VRAM sizing guide shows which models fit on which card.
How the picks were chosen
Every pick is sold new by the manufacturer’s own Amazon store and was in stock on October 19. Each spec comes from the manufacturer’s spec page or, where the maker is silent, an independent review, and every pick was checked against at least one independent source. We left out docks with no manufacturer spec page and docks with a built-in PSU, because none of those (AOOSTAR AG01, AG02 and AG03) is in stock new from the maker right now. The benchmarks are cited from the tests linked above; we did not run our own OCuLink comparison for this guide.
Compare the docks and mini PCs
We split the comparison into two tables with the specs that change the buying decision. For docks that is the connection and the power; for mini PCs it is how the OCuLink port is wired and how much RAM you can add for models that spill out of VRAM.
| Dock | Link | Thunderbolt | PSU | Extras |
|---|---|---|---|---|
| Minisforum DEG1 | OCuLink PCIe 4.0 x4 | No | ATX or SFX, 24-pin, not included | Open frame, any card length |
| Minisforum DEG2 | OCuLink PCIe 4.0 x4 | TB5, 140 W to host | ATX or SFX, 24-pin + 8-pin, not included | M.2 slot and 2.5GbE (over Thunderbolt), hardware toggle |
| AOOSTAR EG02 | OCuLink PCIe 4.0 x4 | TB5, 140 W to host | ATX or SFX, 24-pin + 8-pin, not included | 150 W DC out, mini PC stand |
Two of the three mini PCs below give OCuLink its own port. The UM890 Pro gets it from an adapter in an M.2 slot, which matters if you planned on two SSDs:
| Mini PC | CPU | OCuLink | Max RAM | M.2 slots |
|---|---|---|---|---|
| GMKtec K12 | Ryzen 7 H 255 | Dedicated rear port, PCIe 4.0 x4 | 96GB (spec table) | 3 |
| Minisforum UM890 Pro | Ryzen 9 8945HS | PCIe 4.0 x4 via M.2 adapter | 96GB | 2 (1 with OCuLink in use) |
| Minisforum AI X1 Pro-470 | Ryzen AI 9 HX 470 | Dedicated port, lane count not published | 128GB | 3 |
All three support dual-channel DDR5-5600, so their memory bandwidth for CPU-side layers is the same; the difference is how much RAM fits.
Minisforum DEG1: best OCuLink dock for most builds
Buy this dock if your mini PC already has an OCuLink port. It is a flat open frame with a PCIe x16 slot wired for four PCIe 4.0 lanes, so card length is no problem, and it costs about half as much as the Thunderbolt docks.
You supply an ATX or SFX power supply through the 24-pin header, and the Amazon listing advertises a force power-on mode so the dock can start with the PSU. ChargerLAB measured the bundled cable at about 66 cm. Keep that cable: when Jeff Geerling ran the DEG1 from a Raspberry Pi through an M.2-to-OCuLink adapter, it was picky about generic cables and needed a ReDriver adapter for one card. It does not support hot-plugging, so power the dock before the mini PC.
Best for: mini PCs with a built-in OCuLink port. Pass on it if you also want to plug in a laptop over Thunderbolt. About $110 at the moment.
Minisforum DEG2: best dock for laptops too
The DEG2 adds a Thunderbolt 5 input next to OCuLink, with a hardware switch between them, so the same GPU can serve a mini PC and a laptop. Both paths run at PCIe 4.0 x4, so Thunderbolt 5 does not beat OCuLink on GPU bandwidth.
Inside, the DEG2 packs an M.2 2280 slot, a 2.5GbE port, USB ports and 140 W of power delivery back to the host. Minisforum describes those extras as running over the Thunderbolt 5 link, so a mini PC connected by OCuLink alone should expect only the GPU. Power comes from your own ATX or SFX supply through 24-pin and 8-pin connectors.
Buy it for: a GPU shared between a mini PC and a Thunderbolt laptop. Not worth it for a mini PC that never moves, since the DEG1 gives the same OCuLink link for less than half the price. About $250.
AOOSTAR EG02: best dock that powers the mini PC
The EG02 matches the DEG2’s OCuLink plus Thunderbolt 5 combination for less money and adds a DC output that supplies up to 150 W to a mini PC from the same PSU. A built-in stand holds the mini PC above the dock. AOOSTAR does not publish the DC output’s voltage or plug, so check it against your mini PC’s power input before relying on it, and note the AI X1 Pro-470 has an internal PSU and does not need it.
AOOSTAR says cards with metal brackets three slots wide or more do not fit, so check your card’s bracket before ordering. The Amazon listing is new and had a single review when we checked.
Good fit: a tidy one-cable build with a two-slot GPU. Wrong choice for a three-slot card with a full metal bracket. About $220.
GMKtec K12: best value mini PC with OCuLink
GMKtec gives the K12 a dedicated OCuLink port on the back, wired PCIe 4.0 x4, so all three M.2 slots stay free for SSDs. It ships with 32GB of DDR5-5600 as two 16GB modules, which keeps dual-channel memory bandwidth for anything that runs on the CPU.
No NPU comes with the Ryzen 7 H 255, which does not matter here: llama.cpp has no backend for AMD’s NPU, so the GPU does the work either way. GMKtec’s spec table lists 96GB as the RAM ceiling, though a marketing line on the same page claims 128GB. Plan around 96GB.
Right pick for: a dedicated eGPU host with room for three SSDs. Watch the price: the 32GB listing sells for about 60% more than the $549.99 Wccftech listed for a configured unit in its review. About $880 today.
Minisforum UM890 Pro: most proven OCuLink mini PC
With more than 400 Amazon reviews, the UM890 Pro has the longest track record here. Its Ryzen 9 8945HS is a step up from the K12’s Ryzen 7, and ServeTheHome tested it with a DEG1. OCuLink is wired differently here, though: it comes from an included adapter that sits in one of the two M.2 slots, so with the GPU attached you have one SSD slot left.
RAM tops out at 96GB across two DDR5-5600 slots, and two USB4 ports give a fallback for a Thunderbolt dock. ServeTheHome notes the dock has to be powered on before the mini PC.
Best for: buyers who want the most reviewed hardware and need one SSD. Look elsewhere if you need two NVMe drives with the GPU attached. About $990.
Minisforum AI X1 Pro-470: most RAM for CPU offload
This is the only pick that takes 128GB of RAM, which matters when a model is too big for the GPU and more than 96GB of it has to sit in system memory. Its memory runs at the same DDR5-5600 as the other two, so the gain is capacity, not speed. It also has a dedicated OCuLink port, three M.2 slots and a built-in power supply.
Two things to check. Minisforum does not publish the OCuLink lane count for this model. And Tom’s Hardware found its 32GB review unit shipped with a single memory module, which cut its 3DMark Time Spy score by 44% until it was upgraded to two. The Amazon 32GB configuration may ship the same way, so budget for a matched pair of modules. Partial offload is also the case where link width matters most, and the lane count here is unpublished.
Best for: models that need more than 96GB of system RAM next to the GPU. Too much if your models fit in VRAM, where a K12 does the same job for less. About $1,300.
Check these before you buy
Most OCuLink problems come from the parts around the link, not the link itself. Go through this list before you order:
- Budget for a power supply: none of these docks includes one. Size it for the GPU’s rated power plus headroom.
- Plan on no hot-plug: power the dock first, then boot the mini PC, and shut down before unplugging.
- Find out how the port is wired: a port fed by an M.2 adapter costs you an SSD slot.
- Use the bundled cable: generic OCuLink cables and passive adapters can stop the card from being detected.
- Fit two matched memory modules: single-channel RAM halves the bandwidth for any layers on the CPU.
- Ignore the NPU for this job: llama.cpp and Ollama run the model on the GPU.
If you have not picked the card yet, the GPU guide for local LLMs ranks cards by VRAM per dollar. For a mini PC that runs other services next to the GPU host, see the mini PC guide for Proxmox.
Match the setup to the model size
Start from the models you run. If they fit in your GPU’s VRAM, a K12 or UM890 Pro with a DEG1 is the whole build, and the four-lane link usually costs a few percent: under 5% for a 30B model in Geerling’s test. If you also run a laptop on the same card, swap the DEG1 for a DEG2 or EG02. If your models spill into system RAM, the link and the memory become the limit: expect slower prompt processing than a desktop x16 slot would give, and take the AI X1 Pro-470 with two matched modules only if you need more than 96GB. Past that point, a desktop tower or a unified-memory box like the Ryzen AI Max+ 395 mini PCs is the better tool.
