Using the Digital Twin to design liquid cooling systems that can handle 50 percent higher power densities in AI chips

Meeting standards while accelerating market adoption

The Industrial Technology Research Institute (ITRI), part of the Electronic and Optoelectronic System Research Laboratories (Optoelectronics Labs), is a leading technology research institute in Taiwan, specializing in optoelectronics, electronic packaging, semiconductors and power electronics. The institute is dedicated to driving industry innovation and technological advancements, providing technical support and solutions to businesses in Taiwan and worldwide. To advance in a competitive industry sector, ITRI had several aims to help foster growth. The company saw that, by facilitating technology transfer and application, it could strengthen industry collaboration while advancing Taiwan’s technology sector.

To ensure its products met international standards, while accelerating market adoption, ITRI wanted to enhance power module and system performance, as well as reliability. Additionally, the company needed to drive innovation in electronic packaging technologies, improving thermal management and reliability for semiconductor and optoelectronic products.

To achieve these ambitious goals, ITRI recognized a pressing need for advanced simulation expertise and turned to Flotrend Corporation (Flotrend), a trusted Siemens Digital Industries Software smart expert partner with over 30 years of expertise in advanced thermal and computational fluid dynamics (CFD). Additionally, the company leveraged Simcenter™ STARCCM+ ™ software, which is part of the Siemens Xcelerator business platform of software, hardware and services.

Direct-to-chip cooling of chips on a PCB.

Cooling technology challenges

The rapid rise of generative artificial intelligence (AI) is driving demand for increasingly powerful graphics processing units (GPUs) and AI accelerators across cloud and edge data centers, which traditionally rely on air‑cooled infrastructure. In these conventional data centers, heat load densities typically range from 6 kilowatts (kW) to 10 kW per rack, levels that standard air‑cooling systems were designed to manage.

However, AI factories – facilities purpose‑built for large‑scale AI training and inference – operate on a completely different thermal scale. Due to the exponential growth in performance and thermal design power (TDP) of AI‑dedicated chips, AI factories routinely reach 100 kW per rack or more, far exceeding what any air-cooling architecture can accommodate. Unlike traditional data centers optimized for general compute loads, AI factories concentrate extreme, continuous and highly localized heat generation within dense accelerator clusters, creating far more aggressive cooling demands.

Edge deployments, including factories, retail locations and telecom towers, have their own constraints, requiring compact, efficient cooling solutions like those AI factories require. As a result, designers across these environments are turning to liquid cooling to achieve the thermal performance and energy efficiency that modern AI workloads require. Since liquids have higher thermal conductivity and heat capacity, liquid cooling is more efficient than traditional air cooling. Liquid cooling allows for more effective heat absorption and heat transfer, which consequently enables direct, targeted cooling of hot critical components, and significantly reduces energy consumption for the entire AI factory.

Liquid cooling not only handles higher heat loads but also improves power usage effectiveness (PUE), helping AI factories meet sustainability goals.

Among the leading cooling technologies are direct-to-chip and immersion cooling. Although single-phase fluids are common today, two-phase cooling offers superior heat transfer by leveraging phase change. However, it introduces engineering and operational complexities, such as fluid containment, system sealing and maintenance constraints. Coolant selection also requires careful consideration for performance, cost and environmental impact.

Immersion cooling. Image courtesy of Asperitas, which is used for general representation of immersion cooling. Asperitas is not directly involved in this project.

Maximizing cooling using advanced multiphysics CFD

ITRI recognized the growing demand and inherent limitations of traditional cooling methods. The company was also looking to help its commercial partners maintain their technological leadership. With these factors in mind, ITRI realized they needed to accelerate their research into liquid immersion cooling by leveraging numerical simulation prior to physical testing. Their aim was to accelerate development cycles and improve thermal management in highpower systems. However, they found that experimenting with liquid cooling, especially with two-phase boiling flows, was complex, costly and time-consuming, posing a barrier to faster development.

Seeking accurate and efficient methods for virtual testing and evaluation, ITRI partnered with Flotrend Corporation and leveraged Simcenter STAR-CCM+ for the numerical simulation of the immersed servers, assessing the thermal performance of the design in single and multiphase flow simulations.

Simcenter STAR-CCM+ is a general-purpose CFD software that supports multiphysics simulations, applicable to thermal, fluid dynamics, solid mechanics, chemistry and electromagnetic analyses. It has numerous user-selectable physics models, which facilitated ITRI’s studies into two-phase heat transfer.

“Leveraging Simcenter STAR CCM+ advanced multiphysics capabilities allows us to virtually validate cooling designs, extending confidently into niche areas such as two-phase boiling heat transfer,” says Allen Liu, PhD and principal engineer at ITRI.

Flotrend also offered project-based, advanced customized training and realtime assistance. Their experts provided extensive advice in choosing the appropriate physics models suitable for two-phase boiling flows, ultimately settling on the hybrid multiphase flow model as the most appropriate for the application. This hands-on guidance was instrumental in ITRI’s successful implementation.

With Flotrend’s support, ITRI could rapidly quantify the differences between single and two-phase immersion cooling for a reference printed circuit board (PCB) with four chips. The company considered various liquids during the study, including water and Fluorinert FC-40 and FC-72 dielectric fluids from 3M, all of which had different boiling points.

Since water and FC-40 had higher boiling points compared with FC-72, the virtual tests with these two liquids were in the single-phase regime. The other variables changed in the simulations were the chips’ power dissipations on the PCB. Each of these were made to dissipate 50 watts (W), 100 W and 150.

The objective was to quantify the maximum power density of the chips for the given single or two-phase flow regime. This figure will help designers determine cooling requirements such as coolant selection, design targeted cooling for areas with large power densities, prevent hot spots and increase the reliability and lifespan of the various server components.

In this case, the maximum power density for single-phase cooling was 11.1 watts per square centimeter (W/cm2), without the chips exceeding their maximum allowable temperatures. With two-phase cooling, the cooling efficiency increased by 50 percent, as the maximum power density was found to be 16.67 W/cm2.

“Using the intuitive interface and well-organized physics models within Simcenter STAR-CCM+ made complex two-phase cooling simulations straightforward to set up,” says Liu. “The software’s ease of use lets us focus on innovation rather than tools.”

“Liquid cooling technology not only improves thermal efficiency but also aligns with green AI factory standards, helping businesses achieve sustainability goals,” says Debby Chang, project manager at Flotrend.

Single-phase liquid cooling simulation setup.

Unlocking collaborative research and results

By combining multiphysics CFD, advanced thermal analysis and Digital Twin methodologies, ITRI automated simulation workflows, reduced reliance on testing and performed fast, cost-effective exploration of coolant types, power loads and immersion-cooling configurations. This accelerated development cycles and reduced technical and financial risk.

“These operational improvements have streamlined workflows and enhanced the value delivered by ITRI’s engineering teams thanks to the ability in Simcenter to support multicore parallel processing and GPU acceleration, which improves simulation speed and accuracy,” says Liu.

ITRI and Flotrend have achieved significant advances in immersion cooling technology research. Their findings will aid thermal designers in choosing cooling solutions and coolants for AI factory applications, as well as increase the reliability and lifetime of these high-powered AI chips AI factories are deploying or will deploy. Additionally,
companies can optimize the performance of these servers by maintaining low and constant chip temperatures, which will also help maximize their performance without thermal throttling.

“With the use of Simcenter, we successfully validated two-phase cooling technology, which will have a lasting impact on future AI factory thermal management,” says Mike Wang, PhD and CAX general manager at Flotrend.

Using Simcenter STAR-CCM+ numerical simulation technology empowers ITRI to validate cooling designs, significantly reducing research and development (R&D) risks and accelerating innovation.

“Leveraging Simcenter STAR-CCM+ enhanced our thermal design capabilities, enabling us to achieve 50 percent higher power density in AI chips, proving the viability of two-phase immersion liquid cooling for AI factory applications,” says Liu.

Thanks to the success of their collaboration, ITRI and Flotrend have also jointly published the iMPACT 2022 conference paper, “Thermal Performance of Single-Phase and Two-Phase Immersion Cooling in Data Center.“

Two-phase liquid cooling simulation setup.

Looking Ahead

ITRI and Flotrend plan to expand their work for AI servers and high-performance computing (HPC) applications by researching liquid cooling solutions for high-power density systems. Their goal is to enhance the performance of AI training and HPC systems using improved thermal management.

They also intend to develop new coolants and packaging technologies by investigating more efficient dielectric fluids to increase the stability of two-phase cooling. They plan to participate in global efforts to standardize liquid cooling technologies, supporting broader adoption in AI factories.

Single-phase cooling simulation results showing temperature differences between 50 W and 100 W cases.

Two-phase cooling simulation results showing temperature differences between 50 W, 100 W and 150 W cases.