GPU, iGPU, and an OCuLink eGPU at home
This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.
To run a language model with Ollama, VRAM (the memory on the graphics card) often limits how large a model you can run, and how smooth it feels. This guide is mainly about quantized 7B–13B text chat: an assistant, home automation, summaries. Heavy image generation (FLUX, ComfyUI) is a different stack and usually needs more VRAM. Those paths are ComfyUI and FLUX locally and images without a cloud.
Three kinds of setup show up at home. A tower plus an NVIDIA RTX card. A mini PC plus an eGPU over OCuLink, a high-speed PCIe connector for an external graphics card. Or the iGPU or NPU alone, for light models. An iGPU is graphics built into the processor. An NPU (neural processing unit) is a block of the chip set aside for neural-network math. This guide compares them without invented benchmarks. Cross-check VRAM and model size and the mini PC comparison.
Which setup fits
- Tower plus RTX. You want the most VRAM headroom for 8B–13B chat, sometimes images, and you accept noise and bulk.
- Mini PC plus an OCuLink eGPU. You already have a mini PC with a certified OCuLink port, not a plain USB-C hub, and you want an RTX without a tower.
- iGPU or NPU only. Noise and the bill come first. 3B–7B Q4 chat is acceptable, with a longer wait than an RTX.
A text sketch of the three options
[ATX tower + 650 W power supply] ── PCIe x16 ── [RTX 4060/4070, 8–16 GB VRAM]
[Quiet mini PC] ── OCuLink 4i ── [eGPU enclosure + power supply] ── [RTX]
[Mini PC or laptop] ── shared iGPU ── small 3B–7B Q4 models
Comparison
| Setup | Typical VRAM | Living-room noise | Rough hardware budget, October 2026, plus 36 months of electricity | Best for |
|---|---|---|---|---|
| Tower plus RTX | 8–16 GB dedicated | Medium to loud | €900–1,800 (PC plus GPU) | Heavy family use, ComfyUI and FLUX |
| Mini PC plus OCuLink eGPU | Depends on the card you install | External enclosure, plus fans | €700 mini PC + €300–500 enclosure + GPU | An upgrade without a tower |
| iGPU or NPU only | Shared RAM | Low | Included in a Ryzen AI mini PC | 7B chat, trying it out |
CUDA and offload (NVIDIA). CUDA (Compute Unified Device Architecture) is NVIDIA’s software layer that speeds up inference on a GeForce or RTX card. Offload means some layers of the model are computed on the GPU (VRAM) and the rest on the CPU (RAM).
Option 1 — a tower and an NVIDIA RTX
What you gain: the best performance for the money on GPU inference, cooling you can manage, and a used market. What you give up: bulk, power use (electricity budget), and a child-safety problem if the tower sits in the family office (power cords, front ports, an open case).
Living room and children, with a tower
- Fasten the tower, or put it out of reach of younger children. Hide power strips.
- Explain that the AI-server PC should not be switched off hard in the middle of a long generation.
- Prefer a quiet mini PC if the machine lives in a teenager’s bedroom. See noise, 24 hours a day.
Power supply and cables
- Plan on a 650 W power supply or larger for a mid-range RTX, with some headroom.
- Seat the 8-pin PCIe cable fully. Avoid uncertified SATA adapters for the GPU.
- Use a case with airflow from front to back. Dust means heat, and heat means slowdowns.
Example Amazon partner search: NVIDIA RTX cards on Amazon.com. Check the current price (October 2026).
Option 2 — a mini PC and an OCuLink eGPU
Watch out for USB-C. A USB-C hub, or a generic “eGPU” dock, does not provide the PCIe bandwidth of a certified OCuLink port on the maker’s spec sheet. Without documented OCuLink, GPU inference stays unpredictable, or simply unavailable. Do not confuse a Type-C port on the mini PC with OCuLink 4i or 8i.
OCuLink links a compatible mini PC to an eGPU enclosure plugged into the wall. You get an RTX without a tower, at the cost of a proprietary cable and an enclosure that is often loud under load.
Checklist:
- An OCuLink 4i or 8i port documented on the mini PC spec sheet (Minisforum, some Beelink models; check the 2026 sheet).
- An eGPU enclosure with a power supply of at least 550 W for an RTX 4060 or 4070.
- A cable long enough for the room, and current BIOS firmware from the maker.
- Current NVIDIA drivers. See NVIDIA drivers below for Game Ready versus Studio.
Enclosure search: OCuLink eGPU enclosure on Amazon.com. Amazon partner link. Check compatibility and the current price (October 2026).
Option 3 — the iGPU and the NPU
A Ryzen AI APU (a processor with graphics built in) or an Intel Core Ultra shares system RAM. A 7B Q4 model can run without an RTX, more slowly than on a dedicated card. That is the right call when noise and the bill matter more than speed. To set the CPU and GPU split by hand on a small model, LM Studio GPU offload complements Ollama. For a first graphical install, install LM Studio.
How much VRAM for which model?
| Model you want | Suggested VRAM | Guide |
|---|---|---|
| 3B–7B Q4 | 6–8 GB | A first model |
| A comfortable 8B | 8–12 GB | Family assistant |
| 13B or larger, or images | 12–16 GB or more | Images without a cloud |
Software at home
- Install Ollama detects CUDA on NVIDIA.
- Open WebUI for a family interface.
- Whisper benefits from the GPU too.
Cost and noise
Compare the 3-year cost (Ollama versus ChatGPT) before you buy a new RTX. If the machine runs at night, read noise and power of a mini PC left on all day.
Storage and RAM
A GPU does not replace RAM. Keep at least 32 GB for a comfortable 7B–8B (32 vs 64 GB of RAM). The weights still live on an NVMe SSD.
Safety at home
Do not expose Ollama on the internet. Put guests on their own Wi-Fi. Back up the disks that hold the models.
Further reading
- Choose hardware for home AIMatch uses, RAM, and GPU before you buy.
- Mac mini M4Unified memory, and no dedicated RTX.
- VRAM and model sizeSize the GGUF file, and what spills onto the CPU.
- Hardware selectorBudget, noise, and a GPU profile.
Appendix — upgrade decisions
Before you buy an enclosure or a card, check the OCuLink list (certified port, enclosure power supply, drivers) against VRAM and model size and the mini PC comparison. Those two guides cover “I already bought the mini PC” and “how much VRAM for which GGUF,” without repeating the whole hardware matrix here.
| Situation | What to do |
|---|---|
| Existing tower, tight budget | A used 8 GB RTX, then test ollama run on a 7B Q4 |
| Living-room mini PC, not enough VRAM | An eGPU only if the maker’s spec sheet lists a certified OCuLink port |
| Only a laptop, no desktop | Fine for experiments. Avoid a 24-hour server. Prefer a mini PC for a family assistant. |
Short reminders. A Mac mini has no NVIDIA CUDA: Mac mini M4. A NAS usually has no CUDA path for current Ollama: Synology NAS. Run one heavy session at a time (Ollama or ComfyUI) on an 8 GB card.
NVIDIA drivers, in short
| Branch | Who it is for | For Ollama at home |
|---|---|---|
| Game Ready | Games, frequent updates | Fine. Retest ollama run after a large update. |
| Studio | Video and 3D work | Also fine. Pick the branch you will actually keep updated. |
Download: NVIDIA. Neither branch is required for a household language model, as long as the driver is recent.
A quick decision
- Tight budget, 7B chat. Use the iGPU or a mini PC (comparison).
- Images and chat. Use a tower with 12 GB or more of RTX VRAM (images without a cloud).
- Mini PC already bought, not enough VRAM. An eGPU only if the OCuLink port is certified.
- Absolute quiet in the living room. Skip a gaming tower. Prefer a mini PC and quantized models.
Used cards and a gradual upgrade
Many households start with a used RTX in a tower they already have, before they buy a quiet mini PC for 24-hour use. Check the power supply, test LM Studio, then install Ollama, and estimate the bill (kilowatt-hour budget). If you are building an eGPU, use the OCuLink checklist above. A USB-C hub does not replace OCuLink. The PCIe bandwidth is not equivalent, often 4× to 8× less favorable for GPU inference.
Questions to ask before you buy
| Question | If yes | If no |
|---|---|---|
| Do you need regular FLUX images? | Aim for 12 GB or more of VRAM (VRAM and model size). | For 7B–8B chat, 8 GB can be enough in Q4. |
| Will the machine be visible in the living room? | A mini PC, or a tower with real airflow. Noise, 24 hours a day. | An isolated office tower is acceptable. |
| Will several teenagers use Open WebUI at once? | 32 GB of system RAM, plus a useful GPU. 32 vs 64 GB of RAM. | One 8B Q4 on a 32 GB mini PC is often enough. |
| Do you already have an M4 Mac? | No NVIDIA RTX. Mac mini M4. | Stay with the Apple setup, or add a dedicated Linux or Windows mini PC. |
The software stack after the hardware
Once the card or the APU is chosen, go in this order: a first local AI in 15 minutes, then Open WebUI for beginners, then a private family assistant. The GPU also speeds up Whisper for school voice notes. Do not run a large Ollama model and a ComfyUI generation at the same time on an 8 GB card. Quit one before you start the other.
Note. Rough ranges for October 2026. LocAIhome has not lab-tested every card. Check the maker’s spec sheet before you buy.
Common questions
Is an RTX 4060 enough for Ollama?
8 GB of VRAM (video memory) is a good fit for quantized 7B–8B models. Aim for 12–16 GB for larger models, or for models that are less heavily quantized.
Is OCuLink worth it for a mini PC?
It is useful if the case exposes a certified OCuLink port and you accept an external eGPU enclosure with its own power supply. Check the maker’s compatibility notes.
Can I get by with the integrated GPU?
Recent APUs (Ryzen AI, Core Ultra) speed up small GGUF files. A dedicated RTX is still more effective for heavy inference.
Does a USB-C dock replace OCuLink?
Not for reliable language-model inference. Only a certified OCuLink port on the mini PC spec sheet provides the PCIe link an eGPU expects. A normal USB-C hub is not enough.
How loud is a tower left on all day?
Often louder than a mini PC. See the noise and power guide if the machine lives in the living room.
Sources
Narrow it down with the hardware selector · French version of this section