Immersion Cooling Vs. Direct-To-Chip Cooling: What It Means For Your Piping Specifications

Introduction: The Decision That Happens Too Late

Most data center engineers make their cooling technology choice - immersion or direct-to-chip - based on rack density targets and capital budget. The piping specification comes later, often treated as an execution detail rather than a design driver. That sequencing is a mistake.

The fluid system you build around your chosen cooling architecture isn't interchangeable. A piping spec written for direct-to-chip (D2C) cold-plate cooling will fail - literally and operationally - if retrofitted to single-phase immersion, and vice versa. With GPU rack densities now routinely exceeding 50 kW and AI workloads pushing toward 100 kW per rack, getting the pipe spec right at the schematic design phase is no longer optional. A single leak in a D2C system can take down an entire rack of $400,000 in GPU hardware in seconds.

This guide is written for mechanical and plumbing engineers who have already made (or are making) the cooling architecture decision and need to translate that choice into a defensible, ASHRAE-compliant piping specification.

 

 

Direct to chip liquid cooling data center GPU rack CDU

 

1. Two Architectures, Two Completely Different Fluid Loop Logics

Before touching a pipe schedule, understand that immersion cooling and D2C operate on fundamentally different hydraulic principles.

Direct-to-Chip (D2C) circulates liquid - typically deionized water or a water-glycol blend - through cold plates mounted directly on CPUs, GPUs, and ASICs. Heat is absorbed at the chip surface and carried to a Coolant Distribution Unit (CDU), which exchanges it with the facility water system. The result is a dense, pressurized network of small-diameter pipes, manifolds, quick-disconnect fittings, and cold plates, all of which must work in hydraulic balance simultaneously.

Single-phase immersion cooling submerges entire server boards in a dielectric fluid - typically a synthetic hydrocarbon or engineered fluid - inside a tank. Heat transfers from all components to the fluid, which is then pumped through an external heat exchanger. The external pipe network is minimal: essentially two connections per tank (supply and return) running to a CDU or facility cooling loop.

The architectural difference is stark: D2C distributes the hydraulic challenge across dozens of connection points per rack; immersion consolidates it into a simple, low-pressure external loop. Everything downstream in your piping spec flows from this distinction.

 

2. Pipe Material Selection: Where ASHRAE Sets the Boundary

ASHRAE TC 9.9 and TC 9.6 are explicit on material compatibility. For D2C water-based cooling loops, the guidance prohibits zinc, aluminum, and brass containing more than 15% zinc - all of which accelerate galvanic corrosion in the presence of treated water. In practical terms, this rules out standard yellow brass fittings and galvanized steel pipe that remain common in HVAC hydronic systems.

For the Technology Cooling System (TCS) loop - the secondary loop serving cold plates directly - 304 or 316 stainless steel is the standard specification. The TCS loop typically runs deionized or highly treated water, which is actually more aggressive toward corrosion than municipal water because it has almost no buffering capacity. An isolation valve at each rack, rated for the system's operating pressure and manufactured from 316 stainless, is required both for routine maintenance and emergency leak isolation.

For immersion cooling, the material challenge shifts entirely. The dielectric fluid is chemically inert toward most metals, but it aggressively attacks elastomers - standard EPDM and Buna-N gaskets used in HVAC applications will swell and degrade within months of contact with hydrocarbon-based dielectric fluids. Piping material itself (stainless, carbon steel) is generally compatible, but every gasket, O-ring, and flexible connection in the external loop must be specified as PTFE or a fluoropolymer material rated for the specific fluid in use.

 

3. Pipe Sizing, Flow Velocity, and Pressure Drop

ASHRAE TC 9.6 sets a maximum fluid velocity of 2.1 m/s for pipe diameters larger than 3 inches to prevent erosion at fittings and bends. Below 3 inches, the practical limit is closer to 1.5–1.8 m/s before noise and erosion risk increase unacceptably. These limits apply to both cooling architectures but create very different sizing outcomes.

In D2C systems, flow rate is dictated by the thermal load of each cold plate. As chip TDP (Thermal Design Power) increases - modern H100 GPUs operate at 700W TDP, with next-generation accelerators projected at 1,000W+ - the flow demand per cold plate rises proportionally. A 50 kW rack with D2C cooling may require 8–12 liters per minute of coolant at the rack manifold, distributed across 8 to 16 individual cold plates. This drives the need for careful hydraulic balancing: cold plates in parallel must share flow equitably, or the hottest chips - typically those at the end of the manifold branch - will be starved of cooling capacity.

In immersion systems, the external pipe loop carries the full tank heat load, but at much lower velocity and pressure because the driving mechanism inside the tank is convection and low-speed pumping, not forced high-pressure flow. A 250 kW immersion tank typically requires only 20–30 liters per minute through the external heat exchanger loop, at system pressures rarely exceeding 2–3 bar.

The overhead versus underfloor routing decision also affects sizing. ASHRAE guidance notes that underfloor pipe runs must account for insulation thickness and pipe crossover clearance - adding 50mm to pipe diameter when accounting for insulation wrap is a common planning error that creates coordination conflicts with raised-floor tile layouts.

 

4. Isolation Valves, Strainers, and Leak Control: The Specification Detail That Protects Uptime

This is where piping specifications directly determine operational resilience, and where most specification documents are underwritten.

For D2C systems, each rack must have an isolation valve on both the supply and return legs, enabling a single rack to be isolated and drained for maintenance without shutting down adjacent racks. The preferred valve type for this application is a 3 Piece Threaded Ball Valve in 304 or 316 stainless steel. The three-piece body design is critical here: unlike one-piece or two-piece valves, it can be disassembled in place - without cutting the pipe or removing the valve body from the line - to replace seals or clean the seat. In a live data center environment where minimizing rack downtime is paramount, this in-line serviceability is a meaningful operational advantage, not a specification luxury.

Upstream of each rack manifold, a Stainless Steel Y Strainer is required to protect cold plate micro-channels from particulate contamination. Cold plate geometries for high-performance GPUs use channel widths as narrow as 0.5–1.0mm; a 200-micron particle can partially block a channel and create a localized hot spot that thermal monitoring may not catch until chip throttling begins. Specify strainers with a mesh rating of 50 microns or finer on the TCS loop, per ASHRAE TC 9.9 water quality guidelines.

On the return leg, a Stainless Steel Check Valve prevents reverse flow through the CDU heat exchanger when a rack is isolated or a pump trips - a backflow event that can introduce warm fluid into the supply header and elevate coolant temperatures across the entire row before controls respond.

For server-to-manifold connections, specify dry-break quick-disconnect fittings with a maximum allowable spillage of 1 cubic centimeter per disconnect cycle, in line with OCP (Open Compute Project) connection standards. Flexible hose segments at these connections must be secured with additional clamps at 150mm intervals to prevent sag-induced pressure drop and stress on fitting threads.

For immersion systems, the isolation specification is simpler but not trivial. Each tank needs supply and return isolation valves sized for the full tank flow, plus a sample port for fluid quality monitoring. Specify the external loop connections in a material compatible with the specific dielectric fluid - confirm compatibility with the fluid supplier in writing before finalizing the specification, as proprietary fluid formulations vary.

Leak detection cable should be routed along the full length of all D2C supply piping runs, positioned at the lowest accessible point of each pipe tray or cable tray. For immersion systems, point-sensor leak detectors at each tank base and at the CDU are sufficient, given that external pipe connections are few and high-visibility.

 

 

Leadtek 3 Piece Threaded Ball Valve

 

5. A Practical Specification Checklist Before You Issue for Construction

Before your piping specification leaves the engineering team, verify the following against your chosen cooling architecture:

Material compliance: All wetted metals confirmed free of zinc and aluminum alloys (D2C); all elastomers confirmed compatible with dielectric fluid chemistry (immersion)

Velocity check: Maximum 2.1 m/s at pipes ≥3" diameter, per ASHRAE TC 9.6

Isolation valves: One 3 Piece Threaded Ball Valve per rack, supply and return, with in-line serviceability (D2C)

Strainer rating: ≤50 micron mesh on TCS supply, upstream of each rack manifold (D2C)

Check valve placement: CDU return leg, each rack return branch (D2C)

Quick-disconnect spec: Dry-break, ≤1cc spillage, OCP-compliant (D2C)

Leak detection: Continuous cable on all supply runs (D2C); point sensors at tank base and CDU (immersion)

Underfloor clearance: Pipe OD plus insulation plus 50mm crossover clearance confirmed in floor void model

Water treatment plan: Site-specific chemistry program engaged before system fill, with ongoing monitoring protocol specified

 

Conclusion: Specify the Pipe System to Match the Cooling Architecture, Not the Other Way Around

The cooling technology decision and the piping specification are not sequential steps - they are a single integrated engineering decision. The choice between immersion and D2C cooling sets the hydraulic logic, material constraints, isolation strategy, and leak risk profile of your entire fluid system. Trying to adapt a generic HVAC hydronic pipe spec to either architecture will produce a system that underperforms operationally and creates unnecessary maintenance exposure.

Engage your piping specification at schematic design, alongside the cooling architecture selection. Use ASHRAE TC 9.6 and TC 9.9 as your baseline compliance framework, and verify every material selection against the actual fluid chemistry in your system - not a generic "water" assumption. The 30 minutes spent getting the isolation valve type and strainer mesh rating right at the specification stage will save considerably more than 30 minutes in future downtime.

 

FAQ

Q: Can I use standard HVAC carbon steel pipe for a direct-to-chip cooling loop?

Carbon steel is used in facility-side (FWS) chilled water loops, but it is not appropriate for the Technology Cooling System (TCS) loop that directly serves cold plates. The TCS loop runs highly purified water with low ionic content, which is more corrosive to carbon steel than treated HVAC water. Specify 304 or 316 stainless steel for all TCS loop piping and fittings.

Q: Why does my CDU manufacturer specify a different pipe connection size than my cold plate vendor?

CDUs and cold plates are often sourced from different manufacturers and designed to different internal flow resistance assumptions. The connection size mismatch is a hydraulic balancing problem: if the CDU expects 10 L/min at 2 bar and the cold plates collectively require 14 L/min to maintain chip junction temperatures, neither vendor's spec is wrong - the system design hasn't reconciled them. A hydraulic model of the full TCS loop, from CDU outlet to each cold plate and back, should be completed before finalizing any connection sizing.

Q: How often should Y strainers be cleaned on a D2C cooling loop?

During the first 90 days of operation, inspect strainer baskets monthly - new piping systems shed particulate as pipe coatings cure and fittings settle. After the break-in period, quarterly inspection is typically sufficient for closed-loop systems with ongoing water quality monitoring. If your CDU incorporates a differential pressure sensor across the strainer, set an alarm at 0.3 bar delta-P as the clean-basket baseline.

Q: Is immersion cooling piping simpler to specify than D2C?

The external pipe network for immersion cooling is significantly simpler - fewer connection points, lower pressure, and no rack-level manifold distribution. However, the material specification is more complex because dielectric fluid compatibility must be verified for every elastomer, gasket, and flexible connector in the system. The failure mode of getting material compatibility wrong in an immersion system is slow and hard to detect: seal degradation over 12–18 months, followed by gradual fluid loss and contamination, rather than the immediate and visible leak event that D2C systems produce.

Q: What is the minimum floor void height required for D2C cooling pipe runs under a raised floor?

ASHRAE guidance does not set a universal minimum, as it depends on pipe diameter, insulation specification, and whether pipes cross each other. A practical planning rule is to allow for the largest pipe OD in the run, plus 75mm of insulation on all sides, plus 50mm clearance above for crossover pipes, plus 25mm below for supports. For a 4-inch supply main with 25mm insulation, this yields a minimum floor void of approximately 300mm - larger than the 250mm void common in legacy raised-floor data centers, which is a retrofit constraint worth identifying early.