Each new GPU generation pushes more power through a similar footprint. Most liquid cooling still uses a cold plate bonded to the chip, which adds thermal layers between the silicon and the coolant.
🔥
Rising heat density
More watts through the same area every generation.
🧱
Extra thermal layers
Cold plates and interface materials add resistance.
🏭
Infrastructure strain
Data centres need cooling that scales without more complexity.
01 / InfrastructureAI compute at scaleDense server racks concentrate power and heat.
02 / SiliconCompact electronicsHeat must leave a small, high-power footprint.
03 / HardwareGPU hardwareCooling architecture sits at the heart of performance.
Illustrative real-world photographs for context only; these are not GpuArc hardware or test results.
02The idea
Spray the chip. Let gravity do the drainage.
The GPU is mounted face-down in a sealed chamber. A nozzle sprays coolant directly onto the die. Heated coolant runs down the walls into a small sump, then loops back through a pump and chiller.
GpuArc concept schematic — not a built or tested prototype.
Real-world referenceFrom concept to bench testingA heater-simulator rig could help test flow and heat transfer before risking a GPU.
Validation pathwayMeasure, simulate, iterateCFD, spray coverage and electrical protection remain open engineering questions.
These photographs illustrate the engineering and validation environment, not an existing GpuArc prototype.
01
No bonded cold plate
Coolant meets the die directly, removing a thermal layer.
02
Gravity-assisted drain
Spent coolant leaves the chip surface by gravity.
03
External chiller loop
Heat is rejected away from the chip, keeping the chamber compact.
03How it works
Three steps in a closed loop
STEP 1
Spray
The nozzle directs coolant onto the face-down die.
STEP 2
Drain
Heated coolant returns along the walls into the sump.
STEP 3
Recirculate
A pump moves it through a chiller and back to the nozzle.
04Calculation
First-Order Thermal Analysis
Concept stage, TRL 1-2.
STATUS Concept-stage. CAD in progress. No physical prototype. No experimental validation yet. All figures below are analytical estimates from first-principles heat transfer, not measured results.
Concept
Gravity-assisted direct spray cooling for AI GPUs. The chip is mounted face-down in a sealed chamber. Coolant is sprayed onto the die, drains by gravity to a sump, and is recirculated through an external chiller. This removes the bonded cold plate and its thermal interface layers.
Baseline assumptions
REFERENCE CHIP
H100-class
700 W, die area ~814 mm² (8.14 cm²), heat flux ~86 W/cm²
COOLANT
Deionized (DI) water
Inlet 45 °C, coolant rise 10 K
LIMITS
Max junction 85 °C
Pump efficiency 50%, spray pressure 5 bar
Results (analytical estimates)
Quantity
Result
Coolant flow (700 W, 10 K rise)
~1.0 L/min per GPU
Pump power (5 bar)
~17 W per GPU (~2.4% of chip power)
Rack-level flow (100 kW, 10 K)
~144 L/min
Required spray heat transfer coefficient (h) to hold 85 °C
Ppump = V̇ · ΔP / η = (1.67×10⁻⁵ m³/s × 5×10⁵ Pa) / 0.5 ≈ 17 W
hreq = Q / (A · ΔT) = 700 / (8.14×10⁻⁴ m² × 40 K) ≈ 21,500 W/m²K
Estimated heat removal per die
Q = h × A × ΔT, with ΔT = 40 K. The orange line marks the 700 W baseline chip power.
h = 20,000 W/m²K~650 W
700 W
h = 50,000 W/m²K~1,600 W
700 W
h = 100,000 W/m²K~3,200 W
700 W
Realistic target range: 700 - 1,600 W per die, dependent on achieved h.
HEAT REMOVAL ESTIMATOR
Slide the spray heat transfer coefficient. Area = 8.14 cm², ΔT = 40 K.
h =50,000W/m²K
1,628W
Above the 700 W baseline (on paper)
Key engineering challenges (open)
Electrical protection: DI water contacting a powered die requires a protective barrier or coating, which adds thermal resistance and must be validated.
Spray coverage: achieving h ≥ ~21,500 W/m²K with uniform spray coverage across the die.
Pump penalty: pump pressure and power penalty in direct-to-chip spray configurations.
Die and lid handling: exposing the die for direct spray.
Next steps
01
Complete CAD
Finish the SolidWorks model.
02
CFD / flow simulation
Validate h and spray coverage.
03
Prototype and test
Use a heater-simulator die.
NOTE Figures are first-order estimates for concept evaluation. They are not validated performance claims.
05Prior art
Built on a known approach, with a different goal
Spray cooling is not new. The closest prior art is the ISR/SprayCool patent (US 6,889,515), with a similar chip-down, spray-up, gravity-drain geometry. GpuArc revisits it for modern AI GPU power levels.
PATENT EXPIRED
PATENT US 6,889,515 B2 · "Spray cooling system"
ORIGINAL ASSIGNEE Isothermal Systems Research Inc. (now Parker Intangibles LLC)
FILED 14 Feb 2003 · GRANTED 10 May 2005
EXPIRED12 Nov 2022(anticipated expiry date, status "Expired - Lifetime")
Status as listed by Google Patents; it is not a legal conclusion. This one expired patent does not clear later spray-cooling patents, and no freedom-to-operate check has been done for GpuArc.
Approach
How it differs
Cold plate liquid cooling
Coolant flows through a plate bonded to the chip. GpuArc removes the plate.
Current commercial solutions
JetCool, CoolIT, Motivair and Accelsius still use plate-based designs.
Classic spray cooling PATENT EXPIRED
Same core geometry. GpuArc targets current AI accelerators with a modern external loop.
06Honest status
Where the project actually stands
Stage: TRL 1-2. Concept defined, CAD in early progress.
Not done yet: no physical prototype, no validated thermal data.
Open challenge: pump pressure in direct-to-chip setups, raised by an industry expert.
Next: finish CAD, run flow and thermal simulation, build a bench prototype.
07Collaboration
Let's build this together
GpuArc is an idea I conceived independently as a student. I am not trying to run it alone. The goal is to find the right partners to test it properly, improve it, and find out honestly whether it works.
🤝
Who I want to work with
Cooling specialistsChip and hardware makersHyperscalersAI infrastructure teamsUniversity labsThermal and CFD engineersMentors
🧩
Ways to collaborate
Co-development: refine the architecture together.
Lab partnership: bench testing and a first prototype.
Simulation support: CFD and thermal validation.
Licensing: if the concept proves out.
Technical feedback: especially on pump pressure, electrical protection and coolant choice.
Institutional backing: research or funding routes through my college.
🎯
What I bring
A clearly defined concept with a documented pitch deck.
A first-order thermal analysis with stated assumptions.
Prior-art research and honest positioning against current solutions.
CAD modeling in progress.
Time, energy, and willingness to learn fast and iterate.
🧭
How I work
Transparent: I state clearly what is proven and what is not.
Open to critique: hard technical feedback is welcome.
Credit-fair: contributions and credit are agreed upfront.
08Get in touch
Interested? Reach out.
I'm Abhay, a Mechanical Engineering student (2nd year). If you work in cooling, chips, or AI infrastructure, even a short reply with feedback helps.
