Building an AI PC
What this article covers
- How to define your budget and requirements.
- How to match components together.
- Recommended configurations for entry-level, mid-range, and high-end builds.
- Key dependencies between CPU, GPU, RAM, and motherboard.
- Common mistakes when building.
Introduction: Building an AI PC
An AI PC isn’t just a fast gaming machine. Depending on your intended use, different components will dominate. For local AI work with Ollama, the GPU is usually the most critical factor, followed by RAM and CPU. However, if you plan to run large models, CPU inference, or long-running services, you’ll also need plenty of RAM, a quality motherboard, a robust power supply, and good cooling.
This article shows you how to assemble an AI PC from defining requirements all the way to finalizing your parts list.
Key terms
- Budget frame: Available money for the entire system.
- Requirement profile: What tasks the system should handle.
- Bottleneck: A constraint that limits overall performance.
- Upgrade path: Opportunities for future expansion.
- TDP: Heat output and power consumption.
- Compatibility: Do the components work together?
- ROI: Return on investment, benefit relative to cost.
Step 1: Define your requirements
Before purchasing, clarify these questions:
- What model sizes do you want to run?
- Will you primarily use GPU or CPU?
- How many models in memory at the same time?
- Do you need multi-GPU support?
- Does the PC need to be quiet, or is noise acceptable?
- Will you upgrade it later?
- Is this a workstation PC or a server in the basement?
Step 2: Set your budget
The GPU typically consumes the largest portion of your budget. Here’s a rough breakdown:
| Component | Percentage of budget |
|---|---|
| GPU | 40 to 60 percent |
| CPU | 10 to 20 percent |
| RAM | 10 to 15 percent |
| Motherboard | 8 to 12 percent |
| Power supply | 5 to 10 percent |
| Case | 5 to 10 percent |
| Cooling | 3 to 8 percent |
| SSD | 3 to 8 percent |
Step 3: Choose your GPU
The GPU determines which models will run smoothly.
Recommendations
| Level | GPU | VRAM |
|---|---|---|
| Entry-level, 7B models | RTX 3060 12 GB, RTX 4060 Ti 16 GB | 12 to 16 GB |
| Mid-range, 13B/14B | RTX 4070 Ti Super 16 GB, RTX 3090 | 16 to 24 GB |
| High-end, 30B/70B | RTX 4090 24 GB, two RTX 3090 | 24 to 48 GB |
| Workstation | RTX 6000 Ada, A100 | 48 GB and more |
More VRAM lets you run larger models or use higher quantization levels.
Step 4: Choose your CPU
For pure GPU inference, a modern 6 to 8-core processor with AVX2 is sufficient. If you plan to do CPU inference, run a multipurpose server, or handle many parallel tasks, aim for 12 to 16 cores.
Step 5: Size your RAM
| Model size | Recommended RAM |
|---|---|
| 7B | 16 to 32 GB |
| 13B/14B | 32 to 64 GB |
| 30B | 64 to 128 GB |
| 70B | 128 GB and more |
RAM should run in dual or quad-channel mode to maximize memory bandwidth.
Step 6: Choose your motherboard
The motherboard must match your CPU and GPU. Key considerations:
- Enough PCIe lanes.
- Sufficient DIMM slots.
- Adequate clearance for wide GPUs.
- Quality VRM for stable continuous operation.
- Proper connectors for SSDs and networking.
Step 7: Power supply and case
Your power supply should have at least 20 to 30 percent headroom above your system’s maximum power draw. The case needs space for the GPU, good airflow, and multiple fans. For high-end builds, a full tower often makes sense.
Step 8: Cooling
- Large air coolers or AIO for the CPU.
- Mesh front panel and multiple case fans.
- Regular maintenance and dust filters.
- For multi-GPU setups: Plenty of space and extra ventilation.
Configuration examples
Entry-level AI PC
| Component | Example |
|---|---|
| CPU | Ryzen 5 7600X |
| GPU | RTX 4060 Ti 16 GB |
| RAM | 32 GB DDR5 |
| Motherboard | B650 ATX |
| Power supply | 650 W 80 Plus Gold |
| Case | Mid-tower mesh |
Mid-range AI PC
| Component | Example |
|---|---|
| CPU | Ryzen 7 7700X |
| GPU | RTX 4070 Ti Super 16 GB |
| RAM | 64 GB DDR5 |
| Motherboard | X670 ATX |
| Power supply | 850 W 80 Plus Gold |
| Case | Mid-tower or full-tower |
High-end AI workstation
| Component | Example |
|---|---|
| CPU | Ryzen Threadripper or Xeon |
| GPU | RTX 4090 or two RTX 3090 |
| RAM | 128 GB DDR5 ECC |
| Motherboard | Workstation platform |
| Power supply | 1200 W 80 Plus Platinum |
| Case | Full-tower or rack |
Verify compatibility
- CPU socket and motherboard socket match.
- Motherboard form factor fits your case.
- PCIe version and slot length are appropriate.
- RAM type is supported by the motherboard.
- Power supply fits in the case.
- GPU length fits in the case.
- Cooler height fits in the case.
- GPU power connectors are available.
Plan your upgrade path
If you want to expand later, consider:
- Choosing a motherboard with extra RAM slots.
- Buying a power supply with headroom.
- Picking a case that can hold multiple GPUs.
- Selecting a motherboard with plenty of PCIe lanes.
Common mistakes
- GPU too small: Model won’t fit in VRAM.
- Insufficient RAM: Model gets swapped to disk.
- Weak power supply: System crashes under load.
- Case too tight: GPU won’t fit or overheats.
- Budget motherboard: Too few lanes or RAM slots.
- Forgotten cooling: Heavy sustained load generates lots of heat.
- Wrong PCIe version: Card runs slower than expected.
Further reading and resources
- BotServ.de Buying guide: GPUs for local AI
- BotServ.de Buying guide: CPUs for local AI
- BotServ.de Buying guide: RAM for local AI
- BotServ.de Buying guide: Motherboards for local AI
- BotServ.de Buying guide: Power supplies for local AI
- BotServ.de Buying guide: Cases for local AI
- BotServ.de Cooling for local AI
FAQ: Building an AI PC
What’s the most important part of an AI PC? The GPU with as much VRAM as possible.
Do I need a lot of CPU power? Not for pure GPU inference, but yes for CPU inference and server tasks.
Should I buy one or two GPUs? Two consumer GPUs are more complex but scale better when a model wouldn’t fit in one GPU’s VRAM alone.
What’s better: more VRAM or faster memory? For AI work, VRAM capacity is usually more important than bandwidth.
Is DDR5 worth it? Yes for CPU inference, less so for pure GPU inference.
Sources and further reading
- PCPartPicker: https://pcpartpicker.com/
- BeQuiet Power supply calculator: https://www.bequiet.com/en/psucalculator
- Ollama Hardware: https://github.com/ollama/ollama
Summary: Building an AI PC
A solid AI PC starts with clearly defining your requirements and budget. The GPU with as much VRAM as possible is usually the critical factor. RAM, CPU, motherboard, power supply, case, and cooling must be matched accordingly and leave room for future upgrades. If you pay attention to compatibility, airflow, and upgrade paths, you’ll end up with a system that runs Ollama, RAG, and AI agents reliably and efficiently. Before purchasing, it’s worth comparing several configurations and putting the bulk of your budget into GPU and RAM.