Why Two‑Phase Direct‑to‑Chip Cooling is Reaching a Tipping Point
Modelon & University of Maryland Researchers to Present Findings at Upcoming Conference
Artificial intelligence and high-performance computing (HPC) are fundamentally reshaping the way data centers are designed. Power densities continue to rise, thermal margins are tightening, and workloads are becoming more dynamic and less predictable. HGX H100 racks introduced in 2022 consumed around 40–60 kW, while Blackwell racks introduced in 2024 consume around 120–140 kW, roughly three times more than previous AI racks. Next-generation racks expected in 2026 could reach 200–240 kW. For engineering teams, this shift requires a new approach to rack cooling, one that accounts for higher heat loads, faster transients, and system-level behavior earlier in the design process.
To better understand what comes next, we spoke with Dr. Lingnan Lin, Assistant Professor of Mechanical Engineering at the University of Maryland, College Park, and Kagan Sears, a PhD student in Prof. Lin’s lab. Dr. Lin’s research focuses on phase‑change heat and mass transfer, with applications spanning HVAC systems and electronics cooling. Together, their academic–industry collaboration with Modelon explores two‑phase direct‑to‑chip liquid cooling from both a physics and system‑level perspective. Their assessment cuts across boiling physics, transient system behavior, and practical implementation offering a clear view into why this technology matters now, what engineers are learning today, and where the field is headed.
Why two‑phase direct‑to‑chip cooling matters now
As Prof. Lin described during the interview, the core driver is simple, but profound: the thermal problem has changed faster than cooling architectures have.
“If you look at where AI accelerators and next‑generation processors are going, the heat fluxes are no longer something that traditional approaches were designed for,” Lin explained.
Two‑phase direct‑to‑chip cooling leverages boiling directly at the chip interface, absorbing heat through latent heat rather than sensible heat alone. For AI and HPC workloads with sharp power ramps, localized hotspots, and sustained high utilization, this provides a path to higher heat removal in a smaller footprint.
But neither Lin nor Sears framed two‑phase cooling as a silver bullet. Instead, they emphasized that its promise only becomes real when it is treated as part of a system, not just a superior cold plate.
Why transient system simulation is essential—not optional
A recurring message throughout the conversation was that steady‑state thinking breaks down for two‑phase systems.
“We’re not just interested in whether boiling occurs,” Sears noted. “We want to understand how the system behaves dynamically—as loads change, as flow conditions vary, and as the loop responds over time.”
Boiling initiation, vapor generation, flow instabilities, startup behavior, and control actions all evolve on different timescales. These dynamics directly affect performance, stability, and long‑term reliability. As a result, transient system simulation becomes a design requirement rather than a nice‑to‑have.
Through Modelon’s system‑level simulation capabilities, the team can model
- Interactions between the cold plate, working fluid, pumps, and heat rejection hardware
- Transient operating scenarios such as power ramping and workload variability
- The impact of control strategies on stability and thermal margins
“Without transient modeling, you’re essentially designing blind to the behaviors that matter most,” Sears said.
How continuous feedback loops accelerated research collaboration
Both Lin and Sears emphasized how the collaboration with Modelon changed the pace and depth of their work. Rather than relying solely on long experimental cycles, system simulation enabled rapid iteration at the concept level.
“Being able to adjust assumptions, boundary conditions, and operating scenarios quickly helps us understand which ideas are worth pursuing experimentally,” Sears explained.
From Modelon’s perspective, the value flows in the other direction as well. The academic environment provides a rigorous setting for validation against real two‑phase physics, strengthening confidence in the models and making them more useful for industrial decision‑making.
“There’s a lot of value in closing the loop between simulation and experiment,” Lin remarked. “That’s where real insight comes from.”
This feedback loop—fast iteration paired with physical validation—has allowed the collaboration to move beyond feasibility questions toward system‑level understanding.
Early learnings: what engineers should be paying attention to
While full results will be shared at the upcoming Herrick Conferences at Purdue University, several early observations stood out during the interview:
- System dynamics dominate performance: Small changes in operating conditions can produce qualitatively different behaviors in a two‑phase loop.
- Controls shape outcomes: Stability and efficiency depend as much on control strategy as on component design.
- Simulation reduces experimental risk: Transient models help identify unstable or unsafe regimes before hardware is ever built.
“A lot of the surprises happen at the system level,” Sears noted. “That’s where simulation really earns its value.”
These insights are particularly relevant for engineering managers weighing technology risk and development timelines.
Where two-phase cooling is heading
Looking ahead, Prof. Lin sees the field moving toward integrated design workflows where physics, component design, system architecture, and controls are developed together rather than sequentially.
“The goal isn’t just higher heat flux capability,” Lin said. “It’s predictable, controllable behavior that scales with real systems.”
Validated transient models and digital‑twin‑style approaches are likely to play a growing role, not only in R&D but also in system commissioning and operation. For data centers supporting AI at scale, this predictability may become just as important as raw cooling capacity.
For Modelon, this collaboration reflects a broader strategy: helping engineers move from component‑level optimization to system‑level confidence as data center cooling enters a more dynamic era.
Continue the conversation
The research discussed here will be presented in more technical depth this summer, but the questions facing two-phase direct-to-chip cooling are much bigger than any single study. As rack power densities continue to rise, the industry still needs deeper insight into issues such as refrigerant maldistribution across the rack, working-fluid selection under performance and regulatory constraints, and the implications of two-phase cooling on system architecture, controls, and heat rejection loops.
That is where collaboration between industry and research becomes especially important. Modelon and the University of Maryland team are interested in hearing from data center technologists, cooling system developers, component suppliers, and engineering leaders who are actively evaluating next-generation cooling architectures. If your team is defining requirements, identifying technical barriers, or exploring sponsorship opportunities around two-phase systems and transient simulation, we welcome the conversation.
