Making a Custom GPU Water Block: Cooling Cards With No Product Available
Blocks fit specific board layouts, not GPU models. If no product exists for your card, you either make one or use a universal block. This covers making one.
Consider a universal block first
Before committing, consider a universal block — it cools only the core, with heatsinks and a small fan for VRAM and VRM. It costs a fraction and cannot damage the board if it fails.
Universal blocks fall short on cards with heavy VRAM or VRM heat, or when appearance matters. Only then is a custom block worth it.
Measuring and designing
Design starts from the GPU core centre and the board's mounting holes — the four holes around the core carry the block's pressure.
Measure with calipers, and cross-check with a high-resolution photograph scaled against a known dimension (the PCIe edge connector works well).
The block contacts parts of different heights — core, VRAM, VRM — so the steps must be designed in. Thermal pads absorb small differences, but anything beyond about 1mm should be machined.
Design in CAD; STEP is the usual format to hand to a machine shop.
| Item | Approach | If you get it wrong |
|---|---|---|
| コア接触面 | 微細なフィン（0.3mm前後の溝）で表面積を稼ぐ | 平らな面だけでは冷えない |
| 流路 | 入口をコア直上に当てるジェットインピンジメント形式 | 水が素通りして温度が下がらない |
| 段差 | パッド厚で吸収できるのは概ね±0.5mm | 接触しない箇所ができ、そこが過熱する |
| 固定穴 | 純正クーラーと同じ四隅の穴を使う | 圧力が偏り、コアを割る危険 |
| Oリング溝 | 溝の深さは線径の70〜80% | 浅いと漏れ、深いと潰れて効かない |
Where to have it machined
Without your own CNC, use a machining service. Several accept STEP files online and quote automatically.
Copper (oxygen-free) or nickel-plated copper is standard. Aluminium machines easily but causes galvanic corrosion when mixed with copper. Keep all metals in the loop copper-based.
Make the transparent lid (acrylic or POM) as a separate part, sealed with an O-ring. Acrylic is prone to stress cracking — do not over-tighten.
Assembly and pressure testing
Pressure-test the block before it goes on the board. Skipping this risks a very expensive card.
Plug one port, pressurise through the other, and submerge to check for bubbles. A bicycle pump and gauge works; around 0.5 bar is enough.
After mounting, still run a pump-only leak test for several hours with nothing else powered.
Use the correct pad thickness per component. Too thick lifts the core off; too thin leaves no contact.
Frequently asked questions
What does a custom block cost?
Designing yourself and outsourcing machining, roughly ¥30,000–80,000 in material and machining for a one-off (2026). Iterations cost the same again. A universal block is around ¥10,000.
Can it be 3D printed?
Not the cooling surface — plastics do not conduct heat, so the core contact must be metal. Printed parts work for lids or jigs, but account for water absorption and deformation over time.
Can a design error destroy the GPU?
Yes. Pressure applied away from the mounting holes can crack the die; a wrong step height leaves VRMs uncooled; a leak damages a powered board. Those are the three main risks.