Accelsius has introduced two main thermal testing milestones, demonstrating the unrivaled efficiency and scalability of its proprietary two-phase, direct-to-chip NeuCool™ answer. These breakthroughs tackle crucial cooling challenges for next-generation AI infrastructure and help sustainability targets by means of vital power effectivity enhancements.
In preparation for the subsequent wave of ultra-high-power GPUs, Accelsius examined a thermal check car consultant of a GPU socket. Within the experiment, the chilly plate was heated to 4,500W earlier than reaching the ability restrict of the check infrastructure, not the cooling system. Chilly-plate temperatures remained secure and inside operational limits even at these unprecedented hundreds. This represents the very best cooling capability achieved within the {industry} for direct-to-chip liquid cooling, proving NeuCool has the thermal headroom to help a number of upcoming generations of processors.
In a second check, Accelsius efficiently demonstrated the power of its in-row two-phase CDU to chill a full 250kW rack of AI servers. A four-way H100 server was retrofitted with two-phase chilly plates instantly contacting the switches, CPUs and GPUs, after which put in in a densely configured 250kW rack.
The CDU was subjected to facility water (PG25) at 20°C, 30°C and 40°C, and 375 liters per minute of circulate. Even below most rack load and utilizing 40°C inlet temperatures, the junction temperature of the most popular GPU remained beneath NVIDIA’s thermal throttle restrict (~87°C), highlighting the flexibleness and effectivity of two-phase cooling even in warm-water environments.
The flexibility to function with facility water 6-8°C larger than competing options can translate to over 25% cooling power financial savings whereas enabling extra free cooling hours, relying on the native local weather. An in depth thermal margin was noticed at decrease facility water temperatures, signaling the answer’s scalability to future chip, server and rack architectures.
“We’re exhibiting clients that we will simply meet present efficiency necessities and scale our efficiency to satisfy the wants of the just lately introduced 600kW racks,” mentioned Dr. Richard Bonner, chief technical officer at Accelsius. “Our R&D group has additionally ready us for quickly evolving chip and server architectures, similar to 4,500W TDP sockets and vertically oriented blade servers.”
Accelsius will probably be on the present flooring at Information Middle World in Washington, DC, from April 15–17 (Sales space #524), the place group members will share extra insights from these check campaigns and assist knowledge heart operators consider the trail ahead for AI, HPC and sustainable thermal administration.
Accelsius may even current progressive cold-plate analysis on the OCP EMEA Summit’s Future Applied sciences Symposium in Dublin, Eire, on April 29. The analysis demonstrates the power to chill vertically oriented servers, crucial to the necessities of next-generation NVIDIA Kyber racks.
