13. Budget Build: Entry-Level Under $500
Chapter 13 of 20 · 15 min
The entry-level build targets users running 7B parameter models at reasonable speed. This configuration prioritizes accessibility while delivering capable performance.
Component Selection
| Component | Model | Price |
|---|---|---|
| GPU | RTX 3060 12GB | $280 |
| CPU | Ryzen 5 5600 | $120 |
| Motherboard | B550M | $90 |
| RAM | 2x8GB DDR4-3600 | $60 |
| Storage | 1TB NVMe | $70 |
| PSU | 650W 80+ Gold | $70 |
| Case | Micro-ATX | $50 |
| Total | $740 |
Below $500 total budget requires compromises:
| Component | Budget Option | Savings |
|---|---|---|
| GPU | RTX 3060 12GB used | $80-100 |
| Storage | 512GB NVMe | $40 |
| RAM | 1x16GB (single channel) | $20 |
Performance Expectations
With this build, Llama 3 8B performance:
- INT4 quantization: 22-26 tokens/sec
- Q8 quantization: 16-20 tokens/sec
- FP16: Model does not fit without aggressive optimization
7B models in INT4: Comfortable interactive use 13B models: Requires INT4 quantization, 8-12 tokens/sec
What This Build Cannot Do
- Run 13B models in FP16
- Handle batch inference efficiently
- Support multiple concurrent users
- Run 34B models at acceptable speed
BIOS Configuration
# Recommended BIOS settings for AI workloads
# Memory: XMP/DOCP enabled (DDR4-3600)
# PCIe: Gen 4 (not auto)
# Power: CSM disabled (better UEFI support)
# Storage: NVMe from primary slot (CPU lanes)
Build Notes
- Micro-ATX form factor limits upgrade path
- Budget motherboard may lack PCIe bifurcation
- Single NVMe slot common on B550M boards
- Good thermals require 2-3 case fans
EXERCISE
Search current prices for each component listed. Sum the total and identify the single component where upgrading would provide the most performance benefit for local AI.