🖥️ PochomLab Machine Build Log | RTX 4070 SUPER × 64GB Local Generative AI Setup

■ Overview

This article documents the configuration of the production machine used in PochomLab. It was built as a local environment for generative AI workflows, primarily using Stable Diffusion.

All parts were purchased and assembled between October and November 2025.

■ Purpose

• To document a real-world local generative AI setup
• To record the reasoning behind component selection
• To serve as a baseline for future upgrades and comparisons

■ Assumptions

• Running Stable Diffusion (SDXL) locally
• Supporting LoRA training and ControlNet workflows
• Stable operation under long-running workloads
• Compact Mini ITX form factor

■ Implementation

1. System Specifications

Note: Prices and shops reflect the time of purchase (Oct–Nov 2025).

2. Photos & Load State

Completed Build	Internal Layout
Size comparison with A4 binders	Photo taken during assembly

Idle State (Task Manager)	Under Load (Stable Diffusion)
GPU temperature ~34°C	During Stable Diffusion usage, rises to ~70°C

3. GPU Selection

The most critical component.

Initially, an RTX 4080 (16GB VRAM) was considered. However, due to price and availability constraints, it was not a viable option.

Final choice:

• RTX 4070 SUPER (12GB)

For generative AI workloads:

• SDXL → Fully usable
• ControlNet → Usable with some limitations
• LoRA training → Works with parameter tuning

This resulted in a well-balanced and practical configuration.

4. Case (Mini ITX)

The system was designed with a compact Mini ITX form factor.

Key challenges:

• Heat management
• Maintainability

Final selection prioritized:

• Cooling performance
• Vertical airflow (“chimney” structure)
• Ease of assembly and maintenance

The result is a compact yet accessible internal layout.

5. CPU / Motherboard

• Ryzen 7 7800X3D
• ASUS ROG STRIX B650E-I

Having used AMD platforms for years, the choice was straightforward. This is the first time using a ROG motherboard.

6. Memory (64GB)

• 32GB × 2 = 64GB

Prices were already rising at the time of purchase, so the decision balanced performance and cost.

• EXPO II (6000MHz): not used
• EXPO I (5600MHz): stable operation

For generative AI workflows:

• LoRA training
• Batch generation
• Parallel workloads

64GB provides comfortable headroom.

7. SSD

Selected based on reliability and price-performance balance.

8. Power Supply (850W)

• CORSAIR SF850L

Reasons for selection:

• 850W provides sufficient headroom
• Fully modular for cable management
• Good compatibility with Mini ITX builds

In practice:

• No coil whine (even in winter conditions)
• Quiet operation

Stable under long workloads.

9. Cooling

• DEEPCOOL AK400
• Noctua fans ×3

Designed for long-running workloads such as LoRA training:

• Low noise
• High cooling performance

■ Results

• Stable long-term operation
• Successfully built a local Stable Diffusion environment
• No major issues during assembly

A practical and reliable generative AI workstation was achieved.

■ Notes

• GPU prices increased significantly after purchase
• Memory prices also surged
• Timing of the build was favorable

This domain is highly sensitive to market timing.

■ Postscript

This was my first self-built PC in over a decade, since the Phenom II X3 720 BE era.

Previously, I focused on budget-oriented builds, but this time prioritized stability.

The result:

• Smooth assembly
• Stable operation

A solid foundation for ongoing work in PochomLab.

■ Addendum (Price Situation as of 2026)

This system was originally built for approximately ¥300,000.
However, as of 2026, component prices have changed significantly:

• GPU (RTX 4070 SUPER): over ¥200,000
• Memory (64GB): approximately ¥160,000–¥180,000

Rebuilding an equivalent system today would likely cost around
¥550,000–¥580,000 in total.

PC component prices are highly influenced by market conditions,
highlighting how critical timing is when building a system.