Inside NVIDIA's Secret Supercomputer: Nyx and DLSS Explained

NVIDIA's Nyx Supercomputer

• 🚀 Nyx, NVIDIA's supercomputer at their headquarters, houses 1,192 B200 GPUs, each rated for up to 1200 watts.
• 💡 This powerful system is utilized for diverse applications including AI research, DLSS upscaling for gamers, and Natural Language Processing (NLP) tasks.
• 🏢 The data center is designed for both R&D and NVIDIA's internal production, featuring a reconfigurable layout with high-density water-cooled deployments and air cooling for DGX B200 racks.

Data Center Infrastructure and Cooling

• 🌡️ The deployment is equipped with numerous sensors monitoring temperature, humidity, and air pressure to ensure optimal operating conditions.
• 💨 Cooling is critical, with fresh air blown up through the floor and pushed through DGX units, exiting into a hot aisle, managing up to 14,000 watts of power consumption per unit.
• 🔌 Networking racks are strategically positioned to minimize fiber optic cable length, ensuring signal integrity and reducing strain on the ceiling.
• 🛠️ Cable management is crucial for maintaining airflow and facilitating maintenance, with extra fiber run in each bundle for redundancy.

DGX Rack Design and Power

• ⚡ DGX racks utilize top-mounted PDUs for power distribution, balancing three phases of incoming power.
• 🔌 The system supports high-power connections, such as a 415-volt, 100-amp power plug, and features a design that simplifies the insertion and removal of DGX units.
• ⚙️ A decommissioned engineering unit was torn down to reveal the substantial heat sinks for 1200W GPUs, featuring flat heat pipes for improved airflow and cooling for the Envy Link switching equipment.

DLSS Technology and Development

• 🎮 NVIDIA's GeForce gaming team uses internal compute resources for Deep Learning Super Sampling (DLSS), which renders games at a lower resolution and uses AI to upscale frames.
• 🧠 The development process involves extensive test training runs and iteration on data feeding into DLSS, with teams evaluating new AI innovations.
• 💡 Addressing specific game issues, like graphical glitches in Cyberpunk, can take weeks or months, while more transformative changes, like moving to transformer models, can take over a year.
• 📈 NVIDIA believes DLSS will eventually surpass traditional rendering in quality, citing comparisons to ground truth images where DLSS reconstructions were closer to the ideal than native rendering.
• ⚠️ While DLSS has improved significantly, it still occasionally encounters issues, but the industry widely accepts AI-accelerated image enhancement as the future.