Liquid Cooling in AI Server Racks: How It Works

Liquid cooling in an AI server rack moves heat away from high-power components—CPUs, GPUs, memory, and switches—using water or a dielectric fluid instead of air. You install a cooling distribution unit (CDU), run coolant lines to cold plates or immersion tanks, and reject the heat to a facility loop or radiator. This guide explains the components, the step-by-step setup, and how to keep the system running reliably.

What You’ll Need

• Server rack with liquid-cooled servers or immersion-ready chassis
• Cooling Distribution Unit (CDU) sized for your rack’s heat load
• Coolant (typically propylene glycol/water mix or dielectric fluid)
• Hoses, quick-disconnect fittings, and manifolds (supply and return)
• Leak detection sensors and drip trays
• Facility water loop or dry cooler/radiator for final heat rejection
• Basic tools: wrenches, tubing cutter, pressure gauge, and a multimeter

Step-by-Step

1. Verify rack power and heat load. Check the nameplate power of each server (e.g., 8 GPUs at 700 W each plus CPUs). Sum all components to get total kW. This determines CDU capacity and coolant flow rate.
2. Choose the cooling architecture. For cold plate cooling, each server has a metal plate mounted on the chips. For immersion, servers sit in a tank of dielectric fluid. Cold plates are common for retrofit; immersion is for high-density new builds.
3. Install the CDU. Mount it in the rack or beside it. Connect the facility water supply and return to the CDU’s primary side. The CDU isolates the server coolant (secondary side) from the building loop.
4. Run the manifold lines. Attach supply and return hoses from the CDU to the rack manifold. Use quick-disconnect fittings so you can swap servers without draining the loop.
5. Connect each server. For cold plates, slide the server in and connect the two hoses (inlet and outlet) to the server’s quick disconnects. For immersion, lower the server into the tank and connect power and network cables.
6. Fill the loop with coolant. Use a pump to fill the CDU and lines. Use the recommended mixture (e.g., 30% propylene glycol for freeze protection). Purge air by opening bleed valves at the highest points.
7. Pressure test. Pressurize the secondary loop to the manufacturer’s spec (often 30–60 psi). Hold for 30 minutes and check for drops. Fix any leaks before powering on.
8. Power on and monitor. Start the CDU pump and set the coolant temperature (typically 20–35°C). Boot servers and check inlet/outlet temperatures at each cold plate. Adjust flow rate so all GPUs stay under their thermal limit (e.g., 85°C junction).
9. Commission and log. Record flow rates, pressures, and temperatures. Set up alerts for leaks, high coolant temperature, or low flow.

Common Mistakes

• Under-sizing the CDU: total heat load must be below the CDU’s rated capacity, with headroom.
• Mixing metals (copper and aluminum) without corrosion inhibitors—this causes galvanic corrosion.
• Using tap water—minerals and bacteria clog the loop. Use deionized water with additives.
• Forgetting air purge—air pockets cause hot spots and pump cavitation.
• Not leak-testing before power-on—a drip on a GPU is catastrophic.
• Ignoring the facility loop’s temperature—if the building water is too warm, the CDU can’t reject heat effectively.

Tips

• Use colored coolant or dye to make leak detection easier.
• Place leak sensors under every quick disconnect and along the manifold.
• Keep spare quick disconnects and hose clamps on hand—they wear out.
• Label supply and return lines clearly to avoid cross-connections.
• Monitor coolant conductivity—a rise indicates contamination.
• For immersion, use a fluid with a high flash point (e.g., synthetic esters) and ensure proper fire suppression.

Troubleshooting

• Server overheating: Check coolant flow rate—increase pump speed. Verify that the cold plate is seated properly and thermal paste is applied. Check for air locks by bleeding the loop.
• Leak at a quick disconnect: Replace the O-ring or the entire coupler. Do not try to seal with tape—use proper replacement parts.
• CDU error alarm (low flow): Check for clogged filters, closed valves, or a failing pump. Inspect the strainer and clean it.
• Coolant temperature too high: Increase the facility water flow or lower its temperature. Check that the dry cooler or chiller is functioning.
• Pressure drop across the rack: Likely a blocked hose or a server valve partially closed. Isolate each server to find the obstruction.
• Noisy pump: Cavitation from low fluid level or air in the loop. Top up coolant and bleed air.

Frequently Asked Questions

Is liquid cooling safe for servers?

Yes, when installed correctly. Using dielectric fluid or properly sealed cold plates with leak detection minimizes risk. The main danger is human error during installation—hence the pressure test.

What coolant should I use?

For cold plates, use deionized water with a corrosion inhibitor and biocide, plus propylene glycol if freezing is a concern. For immersion, use a single-phase dielectric fluid like a synthetic ester or PAO.

Can I retrofit an existing air-cooled rack?

Often yes, if the servers have optional cold plate kits. You’ll still need a CDU and hoses. Immersion retrofits require new tanks and are more involved.

How much more efficient is liquid cooling vs air?

Liquid cooling can remove 3–5 times more heat per volume than air. Power usage effectiveness (PUE) can drop from ~1.3 to ~1.05 because fans and compressors use less energy.

What is a CDU?

A Cooling Distribution Unit pumps coolant through the rack and exchanges heat with a facility loop. It provides flow control, temperature regulation, and isolation.

How often should I maintain the system?

Check coolant level and pressure monthly. Replace filters quarterly. Test coolant quality (pH and conductivity) every six months. Replace hoses and O-rings every 2–3 years.

Related

• Prediction Markets vs Hedging: Key Differences
• Delta SkyMiles Platinum Amex: Companion Certificate & Travel Benefits
• Paris Games Biz: Guarded Optimism Amid Turmoil