Feb. 28, 2025 — SoftBank Corp., ZutaCore and Hon Hai Expertise Group (Foxconn) immediately introduced that they carried out ZutaCore’s two-phase direct liquid cooling know-how*1 in an AI server utilizing NVIDIA accelerated computing.
The businesses stated that is the primary implementation*2 of ZutaCore’s two-phase DLC*1 utilizing NVIDIA H200 GPUs. As well as, SoftBank designed and developed a rack-integrated resolution that integrates every part of the server, together with cooling gear with two-phase DLC know-how, on a rack scale, and performed an operational demonstration and efficiency analysis at its information middle in February 2025.
The demonstration outcomes indicated the answer handed NVIDIA’s temperature take a look at (NVQual), thereby confirming the compatibility, stability and reliability of this rack-integrated resolution. The answer additionally achieved pPUE (partial Energy Utilization Effectiveness) of 1.03 (precise measured worth) per rack for cooling effectivity*3.
With the unfold and elevated adoption of AI, demand for AI servers and different computing assets is predicted to develop considerably, additional growing energy consumption at information facilities. On the identical time, decreasing energy consumption from the angle of decreasing carbon dioxide (CO2) emissions has grow to be a urgent challenge worldwide, requiring information facilities to grow to be extra environment friendly, eat much less energy, and introduce modern warmth removing options. Since Could 2024, SoftBank has been collaborating with ZutaCore, a world chief within the growth and enterprise deployment of two-phase DLC know-how, to develop options optimized for low energy consumption at AI information facilities.
NVIDIA H200
SoftBank, ZutaCore and Foxconn will proceed their efforts to realize larger effectivity and decrease energy consumption in AI information facilities utilizing NVIDIA accelerated computing, and collaborate on the event of rack-integrated options optimized with two-phase DLC know-how for sensible use and deployment within the international market.
Options of the rack-integrated resolution embrace:
- Plug & Play resolution
By integrating server parts equivalent to cooling gear geared up with 2-phase DLC know-how, which conventionally needed to be optimized individually, on a rack scale, it’s attainable to offer a “Plug & Play”*4 resolution that can be utilized instantly by merely connecting the gear. This makes it straightforward and easy to put in in AI information facilities.
- 2-phase DLC cooling system
ZutaCore’s two-phase DLC know-how, which cools the chipsets (CPUs, GPUs, reminiscence, and so on.) which are the warmth sources within the server by bringing them into contact with a chilly plate, circulates a particular insulating warmth switch (dielectric) fluid with a low boiling level within the chilly plate and makes use of warmth absorption through the section change from liquid to fuel attributable to the warmth supply. Environment friendly cooling is achieved by way of repeated evaporation and condensation. As well as, the water temperature could be maintained at a better temperature than in typical water-cooled amenities, which improves the warmth transport capability of the water used for cooling and reduces the move fee of water to realize the required cooling impact. This makes it attainable to scale back the strain of the pump that delivers the water for cooling, contributing to decrease energy consumption. Moreover, the usage of insulating warmth switch fluid prevents severe harm to the server within the occasion of a spill of particular fluid.
- AI server with compatibility, stability and reliability
Foxconn developed an AI server based mostly on NVIDIA H200 GPUs, optimized with ZutaCore’s two-phase DLC know-how. This AI server was in a position to exceed goal efficiency figures in NVIDIA’s certification take a look at (NVQual), which determines the compatibility, stability, and reliability of server programs.
- Scalable rack for high-density placement of AI servers
SoftBank developed a rack designed to maximise cooling effectivity and enhance operational effectivity with its two-phase DLC know-how. Along with being designed to be versatile and scalable for high-density placement of the most recent AI servers, the rack has airflow for improved cooling effectivity, and energy provide. As well as, energy provide and wiring are concentrated within the rear of the rack for operational security and robot-friendly operation. The rack can also be appropriate with 21-inch servers and standard 19-inch servers compliant with the ORV3 customary, which is a specification set by the Open Compute Undertaking (OCP), a non-profit group that promotes open-source {hardware} specs and designs. The specs are designed to be deployed in accordance with international requirements.
“As we work to develop one of many largest AI infrastructures in Japan and homegrown large-scale language fashions (LLMs), bettering the power effectivity of information facilities is crucial to speed up AI growth and understand a sustainable society” stated Hironobu Tamba, Vice President and Head of the Knowledge Platform Technique Division at SoftBank Corp. “The operational demonstration of this rack-integrated resolution with optimized two-phase DLC know-how has confirmed secure operation and improved power effectivity of high-density GPU servers with liquid cooling. Sooner or later, we are going to work on the sensible software of this resolution to be used in AI information facilities and AI factories to construct a sustainable AI infrastructure and maximize computational effectivity.”
“ZutaCore is honored to accomplice with international trade leaders SoftBank and Foxconn to introduce the world’s first 2-phase DLC rack integration resolution utilizing NVIDIA accelerated computing for AI servers,” stated Erez Freibach, Co-founder and CEO at ZutaCore. “This groundbreaking collaboration highlights the transformative potential of our waterless DLC cooling know-how to ship unmatched effectivity whereas addressing the rising want for sustainable power options. Collectively, we aren’t simply assembly immediately’s challenges however setting a brand new customary for energy-efficient innovation in next-generation AI information facilities and factories, driving progress whereas safeguarding the planet.”
*1 A know-how that cools the chilly plate on the semiconductor chip (processor) contained in the server by circulating an insulating refrigerant with out water in a two-phase system (liquid and fuel).
*2 As of January 2025, in response to ZutaCore analysis.
*3 In line with ZutaCore analysis. PUE is a measure of cooling effectivity in information facilities. pPUE is a measure of the effectivity of a selected space or facility, equivalent to a server room or module. The nearer the quantity is to 1.0, the extra power environment friendly it’s.
*4 This method that routinely acknowledges peripherals and makes the required settings when peripherals are related to the principle unit of a tool.
