They let me take apart a SUPER COMPUTER

The video dives into Simon Fraser University’s new data center, showcasing a state of the art, liquid cooled supercomputer designed for vast HPC workloads. It begins with a broad overview of the machine’s scale, highlighting the enormous CPU and GPU density, the unique direct die liquid cooling approach, and the potential impact on AI and genomics research across Canada. The presenter then moves into the hardware specifics, revealing the use of 640 Nvidia H100 GPUs, each with 80 GB of VRAM, and the staggering 3.36 TB/s bandwidth delivered by HBM3 memory. He emphasizes that every component, from CPUs and GPUs to RAM and network interfaces, is cooled directly by a liquid-cooling loop, eliminating visible fans and enabling extreme performance. The segment transitions into a tour of the cooling infrastructure, the cooling distribution units, and the evaporative towers that make the data center resilient to heat while maintaining energy efficiency. The discussion wraps with a look at the security measures, the power delivery architecture, and the broader context of who funded and built the facility, underscoring its role in enabling groundbreaking scientific discovery and large-scale data analysis. The middle sections provide a hands-on look at a GPU node and a CPU node, detailing the internal layout, the cooling channels, and the way warm water is extracted from hot components. The host explains the anatomy of the GPU block, including the direct liquid cooling plates and the routing that feeds four GPUs per node, while noting the absence of fans and the tight integration of Lenovo’s system with Nvidia’s GPUs. They discuss the primary and secondary cooling loops, how RAM is cooled, and the clever plumbing that keeps temperatures stable across dense configurations. The narration emphasizes the high power demands and how the cooling loop is designed to manage up to roughly 700 W per GPU, illustrating the engineering trade-offs with tight bends and flow control, all while ensuring reliability through sensors and antimicrobial coolant. The CPU nodes are described as 192-core Zen 5 Epic configurations, with massive RAM and high-speed interconnects, showing how compute resources are distributed across racks and islands to achieve non-blocking bandwidth and sustain performance. Towards the end, the video examines the broader data center ecosystem: the evaporative cooling towers outside, the six-zone security setup, and the telemetry system that feeds data into Kaizen for monitoring and maintenance. The host reflects on the total budget of around 82 million Canadian dollars, the collaboration among vendors like Lenovo, DDN, and Vertiv, and the role of government and academic funding in enabling this petascale system. Viewers get a sense of the scale, from 49 PB of spinning rust to 2 PB of NVMe storage, and the potential to support large scientific endeavours including data from the Large Hadron Collider. The closing notes reiterate the fascination with high-performance computing, liquid cooling, and the future of scientific research powered by such infrastructure, inviting viewers to explore related content and sponsor offers. Overall, the video blends technical depth with accessible commentary, offering a rare, up-close look at a cutting-edge data center that blends hardware, cooling science, and real-world research impact into a cohesive, educational tour.

Topics · science and technology · data centers · high performance computing · cooling technology · artificial intelligence · research infrastructure