NVIDIA Showed Me Their Supercomputer

Quick Overview

NVIDIA's Nyx DGX Supercomputer, located at their Santa Clara headquarters, houses 1,192 NVIDIA Blackwell B200 GPUs totaling over 1.4 PetaWatts of power consumption, utilizing sophisticated Cold Air Containment (CAC) cooling and high-density fiber optic cabling to support AI research and internal production workloads.

Key Points: The NVIDIA Nyx DGX Supercomputer cluster contains 1,192 NVIDIA Blackwell B200 GPUs, with each GPU rated for up to 1,200W, resulting in a total power consumption exceeding 1,430,400W (1.43 PetaWatts). The facility uses Cold Air Containment (CAC) units with perforated floor tiles to actively push cold air up through the racks, which is crucial given the massive heat output. The infrastructure features extensive fiber optic cabling, strategically positioned to minimize length and maintain signal integrity, which is vital for the NVLink fabric networking. NVIDIA's internal production and R&D teams use this supercomputer for tasks ranging from AI research to testing new chip designs, often running workloads continuously. The cooling system is highly engineered, evidenced by the 102 dB reading outside the sealed containment room, necessitating hearing protection (PPE required beyond the point). The presenter held an SXM interface module for the Blackwell GPU, noting its 192GB HBM3e memory and powerful 1200W-1400W thermal design power, which requires the dense liquid/air cooling solutions demonstrated. The video concludes with a discussion of DLSS evolution, noting that newer Transformer models are replacing older CNN/RNN models for better image reconstruction fidelity compared to native rendering (TAA).

Context: The video features a tour and technical deep dive into NVIDIA's Nyx DGX Supercomputer cluster located at their Santa Clara headquarters. The host examines the sheer scale of the system, focusing on its massive computational power, the specialized liquid/air cooling infrastructure required to manage the heat generated by the new Blackwell B200 GPUs, and the advanced networking that connects the compute nodes. The video also contrasts the performance improvements of NVIDIA's DLSS technology (moving from CNN to Transformer models) with traditional rendering techniques.