An official website of the United States government
The rapid growth of AI-driven computing has significantly increased the demand for enhanced cooling technologies to manage escalating chip power densities. While two-phase jet impingement cooling shows promising potential for high-flux dissipation, inherent flow instabilities result in pressure oscillations and localized dryout. This study proposes a two-phase jet impingement cooling architecture integrated with a phase separation membrane. Utilizing water as the working fluid and operating at a sub-ambient pressure, the design was successfully implemented and validated on a commercial high-power NVIDIA V100 GPU for the first time. Experimental results show that the GPU core temperature is well controlled below its maximum operating temperature, achieving a thermal resistance of 0.059 K/W at a GPU power of 300 W. Applying phase separation help mitigate the rise in pressure drop with incasing GPU power, maintaining an average reduction of 15.8%. At a GPU power of 293 W, phase separation reduces both pressure drop magnitude and fluctuation to 5.54 kPa and 0.622 kPa, representing a reduction of 13.5% and 16.0%, respectively, compared to the nonseparation case. This work establishes a scalable, high-efficiency thermal management framework for next-generation AI chips.
more »
« less
This content will become publicly available on December 1, 2026
Integrated Metal-Lidded Microfluidic Cooling on a High-Power AI Chip Using Confined Two-Phase Liquid Jet With Dielectric Fluid R1233zd(E)
Innovative thermal management solutions are required to maintain a lower operating temperature of current lidless packages of AI chips. This study systematically investigates direct-on-chip multiliquid jets cooling over a 2.5-D interposer package using dielectric fluid. A compact metallic and lid-compatible manifold is proposed using 3-D printing technology with alternating impinging and draining nozzles, matching the dimensions of NVIDIA V100 chip. The fabricated manifold is mounted over the stiener and mechanically pressurized with top cover plate and screw arrangement to ensure mechanical robustness and leak proof operation. A dielectric fluid R1233zd(E) is used as a working fluid at a saturation temperature of 37:5 C. Operating GPU temperature is experimentally measured for various flow rates, and power load up to thermal design power of 300 W using embedded sensor within the device. The finding under two-phase operation reveals a significantly lower temperature, measuring lowest thermal resistance of 0.05 K/W with excellent response for step variation in power maps. This enhanced thermal dissipation at the chip level facilitates the compact lidded manifold package and next-generation AI chip.
more »
« less
- Award ID(s):
- 2209776
- PAR ID:
- 10704505
- Publisher / Repository:
- IEEE
- Date Published:
- Journal Name:
- IEEE Transactions on Components, Packaging and Manufacturing Technology
- Volume:
- 15
- Issue:
- 12
- ISSN:
- 2156-3950
- Page Range / eLocation ID:
- 2789 to 2792
- Subject(s) / Keyword(s):
- Advanced packaging, electronics cooling, jet impingement, lidded manifold, two-phase heat transfer.
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
The variety of new electronic packaging technologies has grown significantly over the last 20 years as a result of market demands for higher device performance at lower costs and in less space. Those demands have pushed for heterogeneous packaging, where computer chips with different stack heights are closely packed, creating nonuniform heat flux and temperature and additional challenges for thermal management. Without implementing an appropriate thermal management strategy for heterogeneous packages, large temperature gradients can be observed within the package, which would increase the thermal stresses on the chip and raise reliability issues. To mimic this real-life scenario of such packaging, an experimental setup was designed and built. The design of the new experimental setup consists of four identical 1.2 cm × 1.2 cm ceramic heaters, each of which is connected to a separate power supply and can reach a heat flux of 140 W/cm2. Accordingly, this mock package is capable of delivering different power levels to mimic different multicore microprocessor conditions. To give the heater the ability to move precisely in the x-, y-, and z-directions, each heater is mounted to an XYZ linear stage. Deionized water (DI) was used as the working fluid, and a pin-fin heat sink was used to run the initial steady-state tests on the experimental rig. The tests showed how different flow rates at a constant fluid temperature and input power affect the temperatures of the heaters and the thermohydraulic performance of the heat sink. In addition, a three-dimensional numerical model has been developed and validated with experimental data in terms of heat sink pressure drop and the temperatures of the heaters.more » « less
-
Advancements in electronic device fabrication with increasing integration levels have resulted in very high device densities. This has led to higher power dissipation and heat fluxes, increasing integrated circuit (IC) operating temperature. High and nonuniform heat generation degrades device and system performance. Therefore, thermal management to keep ICs within prescribed temperature limits is an important challenge for reliable and economic performance. Cooling techniques, including liquid coolants and air conditioning (AC), have been utilized to remove heat at the package and system level. However, these techniques must overcome high thermal impedances and require complex integration, while global cooling is generally wasteful, inefficient, and expensive. To improve thermal management, we have developed Si microthermoelectric coolers (μTECs) with areas as small 1E−5 cm^2 that can be integrated on -chip near local hot spots using the standard fabrication processes. While Si μTECs cannot achieve low base temperatures, they can actively pump relatively high heat fluxes directly to a heat sink, thus reducing local temperature increases and allowing targeted rather than global waste heat removal. We demonstrate μTECs that can pump up to 43 W/cm^2 of locally generated excess heat with no increase in chip temperature.more » « less
-
We previously demonstrated how the Sn3Ag4Ti alloy can robustly bond onto silicon via selective laser melting (SLM). By employing this technology, thermal management devices (e.g., micro-channels, vapor chamber evaporators, heat pipes) can be directly printed onto the electronic package (silicon die) without using thermal interface materials. Under immersion two-phase cooling (pool boiling), we compare the performance of three chip cooling methods (conventional heat sink, bare silicon die and additively manufactured metal micro-fins) under high heat flux conditions (100 W/cm2 ). Heat transfer simulations show a significant reduction in the chip temperature for the silicon micro-fins. Reduction of the chip operating temperature or increase in clock speed are some of the advantages of this technology, which results from the elimination of thermal interface materials in the electronic package. Performance and reliability aspects of this technology are discussed through experiments and computational models.more » « less
-
Abstract The rapid growth of digitization and data-driven technologies is driving large-scale deployment of data centers worldwide. As societies become increasingly inference-hungry, the energy consumption of data centers continues to rise at an unprecedented rate, with a significant fraction of this energy being expended on thermal management and cooling infrastructure. Improving cooling efficiency has therefore become a critical challenge for the thermal management community, directly impacting both the sustainability and scalability of future computing systems. In this paper, we demonstrate an energy-efficient cooling solution based on chip-integrated two-phase cooling, which leverages liquid-to-vapor phase-change heat transfer to achieve high heat-flux dissipation at reduced pumping power and thermal resistance. This paper investigates an aggressive cooling architecture utilizing direct-on-die two-phase jet impingement on an NVIDIA Tesla V100 GPU without any thermal enhancement structures. By eliminating thermal interface material (TIM) and impinging the working fluid R1233zd(E), a low-global warming potential (GWP) (<1) refrigerant, directly onto the silicon backside, the primary thermal bottleneck is removed. Experimental results demonstrate a junction-to-coolant thermal resistance of 0.056 °C/W and a theoretical pumping power of only 0.172 W versus the conventional air-cooled cold plate of 0.1709 °C/W and a fan power of 39.9 W. Furthermore, the reliability of the direct-on-die cooling technique was validated through 200 h of continuous, stable operation. This study positions direct jet impingement as a highly efficient, compact, and sustainable solution for next-generation high-performance computing environments.more » « less
