Getting Started with AI PCs
Introduction
You don’t need a server room to start experimenting with local AI. Often a standard consumer PC or used machine is enough for initial work. If you want to buy something new, there are now powerful mini-PCs and workstations designed specifically for AI. The key questions are: how much RAM or VRAM do you have available? Does Ollama run on CPU or GPU? And what’s a reasonable budget for getting started?
AI PC Basics for Beginners
A good entry point for local AI is a PC with at least 16 GB RAM and optionally a dedicated GPU. A GPU with 8 GB VRAM is already enough to run quantized 7B models smoothly. If you only want to use the CPU, you’ll need at least 16 GB RAM, ideally 32 GB, and you should be prepared for slower inference.
Current mini-PCs with AMD Ryzen AI Max+ 395 or Apple’s M4 processors pack a lot of memory into a compact form factor. They’re small, quiet, and powerful enough for larger models. For ambitious workloads, devices like the NVIDIA DGX Spark exist.
Key Terms
- TOPS/NTOPS - Tera Operations Per Second. The unit of measurement for AI compute power. Higher values mean faster model execution.
- APU - A processor with integrated GPU. Examples include Apple Silicon and Ryzen AI Max.
- Unified/Shared Memory - CPU and GPU share the same main memory. Critical for Apple Silicon and Ryzen AI Max.
- Quantization - Reduces a model’s precision to run it with less memory. Enables larger models on smaller hardware.
- VRAM - Memory on a dedicated graphics card. Essential for GPU inference.
- LLM - Large Language Model.
Minimum Requirements
| Component | CPU Setup | Recommended |
|---|---|---|
| RAM | 16 GB | 32 GB |
| GPU | Not required | GTX 1660, RTX 3060 or better |
| VRAM | - | 8 GB or more |
| Storage | 256 GB SSD | 500 GB SSD or more |
A machine with 16 GB RAM and CPU-only setup is sufficient for quick tests. For comfortable work and larger models, a GPU or specialized mini-PC is worth the investment.
Budget-Friendly Entry Options
Option 1: Existing PC
Many PCs with 16 GB RAM can already run small models through Ollama via CPU. This is the cheapest entry point. Speed is limited, but fine for experimentation.
Option 2: Used GPU
A used RTX 3060 with 12 GB VRAM or RTX 3060 Ti with 8 GB VRAM opens up significantly more possibilities. Check that your power supply and case support the card.
Option 3: Apple Silicon Mac
A Mac mini or MacBook Pro with M-series and 16 or 32 GB Unified Memory is a very popular starting point. Unified Memory merges RAM and GPU memory. Many 7B and 13B models run well on it.
Current Mini-PCs for Local AI
Mini-PCs with powerful APUs or dedicated GPU options work particularly well with Ollama, LM Studio, Open WebUI, ComfyUI, and AI agents. They’re compact, quiet, and consume less power than large workstations.
Recommended Selection on Amazon
Mini-PCs for Local AI on Amazon
Bei Amazon ansehenAffiliate-Link: Bei einem Kauf erhalten wir möglicherweise eine Provision.
Ryzen AI Max+ 395
The AMD Ryzen AI Max+ 395 is currently one of the most exciting options for local AI. It combines up to 16 CPU cores with a powerful integrated Radeon 8060S GPU and 128 GB LPDDR5X. That’s enough to run large LLMs, agent workflows, and even image generation locally.
Here are some concrete models available with this processor:
- MINISFORUM MS-S1 Max: Workstation with 64 GB RAM, ideal for large local LLMs. The 128 GB variant wasn’t available at the time of this list.
- GMKtec EVO-X2: Available with 64 GB or 128 GB RAM, 2 TB SSD, USB4, and 2.5 G LAN.
- GMKtec X3: 128 GB LPDDR5X, 2 TB SSD, OCuLink, and card reader.
- Corsair AI Workstation 300: Desktop workstation with 128 GB RAM and 4 TB SSD, good as a stationary AI machine.
- BOSGAME M5: Mini-PC with 128 GB RAM, 2 TB SSD, and Radeon 8060S.
- WEELIAO ONEXStation: 128 GB RAM, RGB lighting, designed for local AI computing.
- Thdeukoty Ryzen AI Max+ 395: 128 GB LPDDR5X, 2 TB SSD, 10 GbE, and WiFi 7.
Prices and availability change quickly. You’ll find current pricing in the Amazon Shop via the card above.
Apple Mac mini M4 Pro
The current Mac mini with M4 Pro and 24 GB Unified Memory is a solid starting point. With 32 GB or more it becomes truly comfortable for local AI. Apple Silicon excels thanks to Unified Memory and efficiency. A Mac mini fits on any desk and runs nearly silent.
NVIDIA DGX Spark and Project Digits
NVIDIA has announced and delivered the DGX Spark, also called Project Digits, a very powerful device for local AI. It’s based on the GB10 Grace Blackwell Superchip and offers up to 128 GB of memory with petaFLOP-scale AI performance. Devices like the ASUS Ascent GX10 or NVIDIA DGX Spark target ambitious users who want to run large models smoothly on-premises.
What About the Ryzen AI Max+ 495?
AMD hasn’t yet released officially available Ryzen AI Max+ 495 processors. When the next generation arrives, it will likely offer even more memory bandwidth and performance. For getting started today, the 395 is a solid choice.
Cost Examples
| Setup | Approximate Cost | Best For |
|---|---|---|
| Existing PC with CPU | 0 EUR | Initial experiments |
| PC + used GPU | 300 to 600 EUR | Solid entry point |
| Apple Silicon Mac mini | 700 to 1,500 EUR | Comfortable entry |
| Ryzen AI Max+ 395 mini-PC | 1,000 to 3,000 EUR | Compact AI workstation |
| NVIDIA DGX Spark / GB10 | 4,000 to 7,500 EUR | Maximum performance in compact form |
Decision Aids and Common Pitfalls
- RAM matters more than raw CPU power. Large models need plenty of memory above all.
- APUs like Ryzen AI Max use shared memory. 128 GB RAM doesn’t mean 128 GB VRAM, but far more than a typical graphics card.
- Mini-PCs are compact but not upgradeable. If you plan for larger models, choose high RAM from the start.
- Apple Silicon is efficient, but not everything runs natively. Some tools and models work better on Linux or Windows.
- Cheap used GPUs remain a good option. If you don’t want to spend much, the RTX 3000 series is still practical.
Further Resources and Links
- 16 GB RAM is the minimum for CPU operation.
- 8 GB VRAM enables quantized 7B models.
- Apple Silicon is popular due to Unified Memory.
- Ryzen AI Max+ 395 and NVIDIA DGX Spark currently offer plenty of power for mini-PCs.
- Learn more about memory in RAM vs. VRAM and RAM and VRAM Requirements.
- Find suitable mini-PCs for local AI in the schutzgeist’s Amazon Shop.
FAQ - Common Questions About AI PCs
Can I use Ollama without a GPU?
Yes, Ollama runs on CPU. However, performance is significantly slower than on a GPU or APU.
Which NVIDIA GPU is best for getting started?
An RTX 3060 with 12 GB VRAM or RTX 4060 Ti with 16 GB VRAM offer good memory for the price. If buying new, aim for 12 or 16 GB VRAM.
Is 16 GB RAM enough on a Mac?
For 7B models, yes. If you want to use 13B or larger models, 32 GB or more Unified Memory is beneficial.
Should I buy a mini-PC or a desktop GPU?
Mini-PCs with Ryzen AI Max or Apple Silicon are compact and quiet. Desktop GPUs often offer more flexibility for later upgrades. For continuously running local AI in your living room, quiet mini-PCs are often the better choice.
Is 128 GB RAM really necessary in a mini-PC?
Not for getting started. For large LLMs, long context windows, and image generation, 64 to 128 GB is a major advantage. This lets 70B models run quantized on the APU sometimes.
Do all AI tools run on Apple Silicon?
Most common tools like Ollama, LM Studio, and ComfyUI run natively or well emulated. Some open-source models are first optimized for CUDA and run better on Linux or Windows with NVIDIA GPUs.
What’s the advantage of Ryzen AI Max+ 395 over a graphics card?
The processor combines CPU, GPU, and up to 128 GB memory in one compact package. It uses less power than a large workstation while remaining powerful enough for large local models.
Is the NVIDIA DGX Spark worth it for beginners?
No. The DGX Spark targets professionals and very ambitious projects. For getting started, a used PC, a cheap GPU, or a Ryzen AI Max mini-PC is enough.