3D-Printed Direct-Water GPU Cooler Drops RTX 2060 Super Load Temperature to 28°C

Tom Hardware has documented TrashBench pumping coolant directly across bare GPU silicon, reaching 28°C under load after several prototypes leaked.

The Australian modder removed the metal cold plate that normally separates coolant from a graphics processor. A 3D-printed chamber instead guided water across the exposed die of an Nvidia GeForce RTX 2060 Super.

That final result beat both the card’s stock cooler and a clamped-on all-in-one liquid cooler. However, the experiment also exposed the immense gap between a working prototype and a cooling system anyone should trust.

The stock configuration reportedly reached 70°C during the comparison. An all-in-one cooler lowered that reading to 36°C, while the direct-water arrangement reached 28°C under the same reported workload.

Those figures make the experiment more than another eccentric PC modification. They illustrate how removing layers from the thermal path can improve heat transfer.

They do not establish a safe upgrade path for gaming PCs. TrashBench encountered leaks, porous printed material, questionable seals, and exposed components before obtaining the headline result.

The real contest is therefore not water versus air. It is minimum thermal resistance versus the reliability delivered by a conventional metal water block.

Tom Hardware Found a 42°C Drop After the Prototype Stopped Leaking

The experiment produced an extraordinary temperature reading only after repeated failures exposed how difficult direct-water cooling is to contain.

TrashBench began with a simple question. If a metal water block transfers heat from silicon into coolant, why keep the metal between them?

A conventional GPU block presses a metal cold plate against the graphics processor. Thermal interface material fills microscopic gaps between those surfaces, while internal fins transfer heat into moving coolant.

TrashBench removed that entire middle stage. His printed chamber placed flowing water in direct contact with the exposed surface of the GPU die.

According to the original cooling experiment, he first measured a nonfunctional GeForce RTX 3060. That card provided a lower-risk platform for fitting and leakage tests.

He coated nearby surface-mounted components with nail polish. The coating served as an improvised barrier against water reaching electrical contacts beside the processor.

The 3D-printed block incorporated washers, gaskets, clamps, and plumbing fittings. Some fittings were melted into the printed structure, while the original cooler’s retention mechanism held the assembly against the card.

The first prototype leaked. TrashBench responded by applying epoxy around the block, then discovered coolant escaping near the tube fittings.

Those fittings received more epoxy. The sequence eventually produced a prototype that appeared to hold water without visible leakage.

A powered GTX 980 then served as the first live test subject. Once that older card survived, TrashBench installed a revised design on the RTX 2060 Super.

The final comparison produced three reported GPU temperatures under load:

• Stock cooling reached 70°C.

• The clamped-on all-in-one cooler reached 36°C.

• Direct water on the die reached 28°C.

That makes the direct arrangement 42°C cooler than stock and 8°C cooler than the AIO result. It also places the final load temperature close to typical room conditions.

These readings remain results from one creator’s experimental setup. They are not independent laboratory validation, and the available reporting does not establish a complete controlled test protocol.

Important variables include ambient temperature, coolant temperature, pump flow, radiator capacity, fan speed, GPU power, clock behavior, and benchmark duration. Each can materially alter the reported temperature.

The result still has value. All three configurations reportedly ran on the same RTX 2060 Super, making the comparison more informative than unrelated measurements from different systems.

TrashBench also used a PCIe riser cable during powered tests. That precaution separated the graphics card from the motherboard, reducing the chance that a leak would immediately reach another expensive component.

This detail captures the experiment’s central conflict. The modification worked well enough to produce an impressive number, yet its test arrangement assumed that another leak remained possible.

Why Removing the Cold Plate Changes the Thermal Path

Direct water contact shortens the route between hot silicon and coolant, but it also removes the engineered surface that makes ordinary water blocks effective.

Heat from a working GPU starts at transistor junctions inside the die. It must travel through silicon, interface material, a cold plate, and finally into coolant.

Every boundary creates some thermal resistance. Thermal resistance describes how strongly a material or interface opposes heat moving through it.

Removing a cold plate and its thermal interface eliminates two boundaries. In principle, coolant can then collect heat closer to its source.

That principle helps explain the 28°C result. It does not mean a flat silicon surface automatically transfers heat better than every metal cold plate.

Commercial blocks contain dense fins or microchannels that create far more wetted surface area. Their geometry also disturbs the flow, helping fresh coolant reach the hot material.

A shallow chamber over a smooth die has less exchange area. Water can also follow an easy route between inlet and outlet, leaving slower regions elsewhere inside the chamber.

TrashBench’s later CPU test reportedly highlighted the importance of chamber size. A larger or poorly shaped cavity can reduce velocity across the die and weaken local heat transfer.

This makes the experiment a fluid-distribution problem as much as a materials experiment. Chamber height, inlet direction, outlet position, flow rate, and pressure all affect performance.

Direct-water cooling also differs from conventional enthusiast “direct-die” cooling. In a typical direct-die system, a metal block still contacts the silicon after a heat spreader is removed.

TrashBench went further. His coolant touched the exposed die instead of flowing through a metal block mounted above it.

The distinction matters because a cold plate performs several jobs. It spreads heat, provides internal surface area, distributes pressure, guides coolant, and creates a durable sealing boundary.

Removing it can reduce one part of the thermal path. It simultaneously transfers all those mechanical and hydraulic responsibilities to the printed chamber and the chip package.

Research on on-chip cooling supports the general value of reducing thermal distance. One peer-reviewed microchannel study describes channels integrated near a chip’s heat source, simplifying a longer chain of interfaces.

However, those systems do not simply place a hobby-grade plastic cap over working electronics. Researchers fabricate controlled channels, bonded structures, sensors, and carefully characterized seals.

TSMC has also reported direct liquid cooling for high-performance packages. Its published work compared lidded cooling with a more direct approach at a specified coolant flow.

The company reported lower junction-to-ambient thermal resistance for direct liquid cooling. That supports the broader mechanism behind TrashBench’s result.

The scale and engineering discipline are completely different. TSMC’s work addresses integrated semiconductor packaging, while TrashBench tested a removable printed enclosure on consumer hardware.

That comparison makes the home experiment more interesting, not more practical. It shows the same physical incentive appearing at opposite ends of the computing market.

The 28°C figure therefore should not be interpreted as proof that metal is unnecessary. It shows that removing interfaces can help when a crude replacement still maintains adequate flow.

A better-designed metal block might narrow the gap. A worse chamber could erase it entirely, regardless of direct coolant contact.

The Real Opponent Is Reliability, Not Cooling Performance

TrashBench’s design beat the tested coolers on temperature, but conventional blocks retain a decisive advantage in predictable containment.

A water-cooling component must do more than survive a benchmark. It must tolerate thousands of heat cycles, pump vibrations, pressure changes, coolant chemistry, and long periods without inspection.

The TrashBench direct-die cooling project had problems before the RTX 2060 Super cooling comparison even began. Its earliest printed chamber leaked and required epoxy around several possible escape paths.

The chosen printing material also mattered. According to the report, some versions leaked around their edges, while other prints were slightly porous.

Porosity presents a difficult failure mode. A joint can look sealed while coolant slowly migrates through microscopic voids in the printed structure.

Material behavior can also change with temperature. A plastic that holds its shape during a short test might creep after repeated exposure to heat and mounting force.

Epoxy can patch visible gaps, but it does not create a qualified consumer cooling product. Its bond depends on surface preparation, cure conditions, geometry, and compatibility with surrounding materials.

The GPU package adds another challenge. A processor die is hard but brittle, and its surrounding components were never designed to form part of a plumbing enclosure.

Mounting pressure must compress a gasket enough to stop liquid. Excessive pressure can damage the die, package, circuit board, or surrounding solder connections.

Too little pressure creates another leak path. Uneven pressure can produce both poor sealing and mechanical stress at different points around the chamber.

A conventional full-cover GPU block solves these problems with a machined cold plate, purpose-built seals, rigid mounting hardware, and defined contact areas. It also cools components beyond the core.

That wider coverage matters. Graphics memory, voltage regulators, and power-delivery components also generate heat under load.

The reported 28°C figure concerns the GPU core. It does not establish how the direct-water chamber managed every other heat source on the card.

A stock cooler usually links those components into a shared thermal assembly. A full-cover water block similarly provides contact areas for the core, memory, and power circuitry.

A small chamber focused on the die needs separate cooling for everything else. Otherwise, a spectacular core reading can coexist with overheated memory or voltage-regulation hardware.

Water conductivity creates another risk. Pure water begins with relatively low conductivity, but it can collect ions and contaminants from fittings, residues, dust, and exposed materials.

Nail polish around surface-mounted components is an inventive precaution. It is not equivalent to a validated conformal coating process across the entire board.

The riser cable also mitigated consequences without solving the leak. It moved the card away from the motherboard, giving escaped coolant fewer immediate targets.

Long-term operation would require leak detection, controlled pressure, compatible coolant, stable seals, and protection for nearby electronics. The design would also need repeatable assembly.

These requirements explain why commercial cooling companies keep the water inside a block. The metal is not present only to conduct heat.

It also forms a durable, machinable barrier between liquid and delicate circuitry. Manufacturers can pressure-test that barrier before it reaches a customer.

TrashBench’s experiment challenges the cold plate’s thermal role while reinforcing its structural role. The modification gained temperature headroom by discarding a mature safety boundary.

That is a fair exchange in a monitored experiment using expendable hardware. It is a poor exchange inside an unattended daily workstation.

The RTX 2060 Super Result Needs More Than One Temperature

A 28°C core reading is compelling, but cooling performance cannot be separated from coolant conditions, card power, and test duration.

Temperature comparisons become meaningful only when the surrounding conditions remain controlled. A lower number can reflect a better thermal interface, colder coolant, lower power, or several effects together.

The source comparison identifies stock, AIO, and direct-water results. It does not provide enough published detail to reconstruct every test variable from the article alone.

That limitation does not invalidate the measurements. It narrows the conclusions readers should draw from them.

The clearest supported conclusion is that TrashBench’s final direct-water setup cooled this RTX 2060 Super better than the two tested configurations.

The experiment does not establish that every direct-water design will beat every GPU block. It also does not quantify performance per unit of flow, pump power, noise, or radiator capacity.

Coolant temperature is especially important. A cooling loop cannot normally hold a GPU below the temperature of the liquid entering its chamber.

TrashBench also experimented with coolant at minus 28°C, according to the report. That extreme test should remain separate from the standard 28°C load result.

Subambient coolant introduces condensation. Moisture from surrounding air can collect on tubing, fittings, the package, or the circuit board when their surfaces fall below the dew point.

Condensation can reach areas that never contact the coolant loop. A perfectly sealed chamber therefore does not eliminate electrical danger during subambient operation.

The experiment also reportedly used an Intel Core i5-7600K. TrashBench selected that processor partly because losing it would not be especially costly.

That choice says little about modern CPU performance, but it helped demonstrate the effect of chamber dimensions. Direct contact cannot compensate for poorly managed flow.

Benchmark duration is another open question. A short run can capture excellent core temperatures before coolant and radiator temperatures reach equilibrium.

A longer test reveals the loop’s sustained capacity. It also increases the opportunity for slow leaks, seal movement, and component temperatures outside the GPU core to become visible.

Clock speed and board power deserve equal attention. Nvidia’s GPU Boost controls allow frequency to respond dynamically to workload and thermal limits.

A colder GPU might hold higher boost clocks and consume different power. Two tests showing different temperatures may therefore represent slightly different operating points.

Useful follow-up data would include inlet temperature, outlet temperature, coolant flow, ambient temperature, GPU power, core clock, hotspot temperature, and memory temperature.

A pressure measurement would help too. It would reveal whether the printed chamber creates excessive restriction or depends on unusually high pump force.

The most convincing comparison would use the same pump, radiator, fans, coolant, power target, and benchmark duration. Only the interface and chamber would change.

Repeated assembly tests would also matter. A system that performs well once but leaks after remounting has little practical value.

The phrase “direct die water cooling” can also cause confusion. Many commercial direct-die products remain carefully machined metal water blocks.

TrashBench’s direct-water design is more radical because the liquid itself replaces the cold plate. Readers should not treat the risks as applying equally to every direct-die product.

For enthusiasts, the correct response is curiosity rather than imitation. The reported performance deserves attention, while the incomplete controls and evident failure modes demand caution.

A Garage Experiment Mirrors the Data Center Cooling Race

The homemade chamber points toward a serious industry goal: moving coolant closer to heat as chip power density rises.

TrashBench built his system around an older consumer GPU, common plumbing hardware, printed plastic, and epoxy. Major chipmakers approach the same thermal principle through semiconductor manufacturing.

Microsoft has tested microfluidic channels etched into the backs of silicon chips. Microfluidics controls small liquid volumes through extremely narrow passages near concentrated heat sources.

The company says its laboratory system removed heat up to three times better than conventional cold plates under certain workloads and configurations.

Microsoft also reported a 65 percent reduction in the maximum temperature rise inside a GPU. Those are company results, not a universal performance guarantee.

Its microfluidic cooling design differs fundamentally from TrashBench’s open chamber. Engineered channels increase surface area and route coolant toward specific hot spots.

This approach addresses a weakness in simply flowing water over flat silicon. Closely spaced channels create more contact area and more controlled local flow.

Researchers can also tailor a channel network around nonuniform heat. Modern processors do not produce identical heat across their entire surface.

Cache, compute units, memory interfaces, and other regions can create different thermal loads. A uniform chamber might overcool one region while underserving another.

Integrated microchannels promise tighter thermal control. They also introduce manufacturing, clogging, inspection, repair, and reliability challenges.

A tiny obstruction can change flow through a narrow channel. Corrosion or deposited material can become significant at microscopic scales.

The coolant must remain compatible with every material it contacts. Pump pressure, seals, channel walls, and package bonds must last through the product’s expected service life.

These constraints explain why direct-to-silicon cooling has developed gradually despite its physical appeal. Better heat transfer alone does not make a deployable system.

Data center operators also evaluate serviceability. Technicians need to replace servers or cooling components without exposing expensive hardware to uncontrolled liquid.

Consumer PCs place similar demands on a smaller scale. Owners expect a graphics card to survive years of movement, dust, temperature swings, and occasional maintenance.

The common challenge is packaging. Both a hobbyist chamber and an integrated microfluidic cooler must bring liquid closer without making the computer less dependable.

The direct water cooling GPU experiment offers a visible demonstration of that tradeoff. Every removed thermal layer also removes material that previously protected, spread, or contained something.

For data centers, solving the tradeoff can unlock denser computing. Warmer facility water could collect useful heat without requiring energy-intensive chilling.

For consumer graphics cards, the incentive is weaker. Existing air and liquid coolers already keep most cards within their designed operating ranges.

Lower temperatures can reduce noise or preserve boost behavior. However, the practical benefit often remains modest compared with the consequences of a leak.

The RTX 2060 Super result matters because it makes thermal resistance tangible. A 42°C reduction from stock is much easier to understand than an abstract engineering coefficient.

Yet the industrial lesson comes from the failures surrounding that number. Future direct-to-silicon systems will succeed only when their containment becomes less remarkable than their performance.

What Would Turn Direct-Water GPU Cooling Into a Credible Design

Three tests will determine whether TrashBench’s idea remains a dramatic prototype or becomes a repeatable cooling architecture.

The first signal is endurance. A revised chamber should complete extended pressure and thermal-cycle testing without epoxy repairs, visible moisture, or changes in mounting force.

A single benchmark proves that coolant moved and the card survived. Hundreds of heating and cooling cycles would say much more about seal stability.

The test should include periodic inspection and pressure monitoring. A slow loss of pressure can reveal leakage before liquid becomes visible around the card.

If a printed chamber survives those tests, the experiment’s reliability argument strengthens. Another leak would reinforce the case for a machined or bonded enclosure.

The second signal is a controlled thermal comparison. TrashBench should test direct water against a purpose-built GPU block using identical loop components and operating conditions.

That comparison needs matched coolant temperature, flow rate, radiator settings, board power, ambient temperature, and benchmark duration.

Core, hotspot, memory, and voltage-regulator temperatures should appear together. Clock speeds and power readings would show whether colder silicon changed the card’s behavior.

If direct water retains a clear advantage under those controls, removing the cold plate deserves further investigation. A much smaller gap would shift credit toward loop conditions.

The third signal is chamber geometry. Follow-up prototypes should vary cavity height, inlet placement, outlet placement, and internal flow structures.

Transparent test sections or flow visualization could reveal stagnant zones. Temperature sensors at the inlet and outlet could quantify how much heat the coolant carries away.

A structured design study would connect TrashBench’s garage experiment with serious microfluidic research. It could show whether simple jet impingement beats a shallow open chamber.

Jet impingement directs a fluid stream at a hot surface, creating strong local heat transfer where the stream lands. Multiple jets might spread that effect across the die.

Internal pins could increase wetted area, but they would complicate printing and cleaning. Narrow passages might also require more pump pressure.

Material choice will remain central. A printed polymer must resist coolant, pressure, temperature, and mechanical creep without allowing fluid through its walls.

A machined transparent top paired with a controlled gasket might offer a safer development platform. It would retain visual inspection while improving dimensional stability.

None of these changes would make the design suitable for casual installation. Exposed silicon, board-level components, and liquid remain an unforgiving combination.

Tom Hardware captured the experiment’s most attractive fact: an RTX 2060 Super reportedly ran at 28°C instead of 70°C under the compared load.

The more important result is the unresolved engineering bill behind that temperature. Every removed layer must be replaced by better flow control, sealing, protection, and validation.

Watch the next TrashBench direct die cooling tests for endurance data, controlled loop comparisons, and redesigned chamber geometry. Those signals will reveal whether 28°C was a repeatable architectural gain or one successful run after several leaks.