Johnson Eung is the senior growth products manager in AI supercomputing at Supermicro, with over 13 years of experience across data centers, spanning cloud infrastructure, high-performance computing, and AI. In this DCD>Talks episode, Eung reflects on the recent AI surge, noting:

“Around 2023 was the inflection point where the world really took notice of AI, and suddenly what I did became interesting to others.”

At Supermicro, Eung brings a holistic approach to AI, focusing not just on launching cutting-edge hardware but on the entire AI stack – from compute to network architecture and energy efficiency.

Modular foundations

Led by the visionary Charles Liang, Supermicro has built a 30-year legacy of modular, high-performance solutions. Known for designing and manufacturing energy-efficient servers, storage systems, and software for enterprise, cloud, AI, and 5G markets, Supermicro has established itself as a leader in the industry.

“We want to make every step of the solution more modular – when we developed our first motherboards, we designed them to work with as many platforms as possible,” says Eung.

Supermicro’s approach ensures that each new product is backward-compatible with its predecessors, creating a foundation for the next generation of building blocks.

Bridging the gap

It’s rare to see a company maintain a consistent vision for over 30 years, yet Supermicro has done just that. While today’s data centers may look similar to those of the past, the demands of AI have created a unique set of challenges. How does Supermicro’s modular building block approach bridge the gap between the compute, power and cooling, and networking in AI data centers?

Eung takes us back to the fundamentals, highlighting recurring challenges in AI data center deployment: insufficient power, inadequate networking, or a lack of design and engineering expertise. He offers perspective, stating that a typical data center rack draws about eight to 13 kilowatts – about the same as charging three electric vehicles (EVs) or powering a single American household. In contrast, today’s AI racks consume up to 160 kilowatts – equivalent to a Whole Foods rooftop covered in solar panels or 45 EVs charging simultaneously.

“Sure, these look like standard servers, but the infrastructure needed to cool and power them at this scale doesn’t exist. So we’ve worked with customers to create this from the ground up, providing clear parameters to ensure it’s done responsibly, safely, and sustainably,” Eung explains.

Liquid cooling as a crucial building block

The trend is clear: performance and power consumption are both increasing dramatically. This rising power equals more heat, and while data centers traditionally use air to cool servers, liquid cooling is far more efficient. According to Eung, Supermicro addresses this challenge with direct-to-chip liquid cooling, where metal plates on CPUs and GPUs efficiently transfer heat to circulating liquid:

“This method saves around 40 percent of the electricity previously used for air conditioning. And as AI continues to grow rapidly, we’re working on even better ways to cool servers efficiently, reduce water usage, and push sustainable, green computing forward.”

The future of AI data center connectivity

As AI becomes increasingly complex, the demand for GPUs soars, driven by the skyrocketing number of tokens needed for realistic video and natural conversations. But these GPUs don’t work alone – they need to communicate like neurons in a human brain, without latency, creating an unprecedented networking challenge in data centers.

To address this, Supermicro employs a fat-tree topology with fiber optic connections, ensuring every GPU has a direct, high-speed link to every other GPU, as Eung contextualizes:

“For example, in a system with 1,000 GPUs, you’d need about 22,400 meters of fiber optic cables – like 224 Olympic runners covering a relay! To make scaling manageable, we provide pre-connected GPU building blocks that simplify deployment, backed by dedicated on-site support. This approach enables us to build what we call “AI factories,” powering the next generation of AI innovation.”

To watch the full DCD>Talks episode, click here.