Skip navigation
This is Nvidia's new $8,000,000 GPU rack. Except it isn't one GPU. It is seventy-two. 🖥️⚡
0:00

This is Nvidia's new $8,000,000 GPU rack. Except it isn't one GPU. It is seventy-two. 🖥️⚡

#nvidia #gpus #verarubin

This is Nvidia's new $8,000,000 GPU rack. Except it isn't one GPU. It is seventy-two. 🖥️⚡ Meet Nvidia Vera Rubin: the next-generation AI supercomputer architecture designed to replace Blackwell NVL72. Here is the exact hardware engineering behind the rack: 1. Transistor Scale: Pull out a single compute tray. Inside sit four Rubin GPU chips. Each Rubin die packs a staggering 336 billion transistors built on cutting-edge 3nm process nodes with next-gen HBM4 memory stacks. 2. The Vera CPU Core: The rack pairs 72 Rubin GPUs with 36 custom Vera CPUs designed in-house by Nvidia. By controlling both host CPU silicon and accelerator GPUs, Nvidia eliminates host-to-device PCIe bottlenecks. 3. The NVLink 6 Passive Spine: How do 72 separate chips act as one coherent giant GPU? At the rear of the rack sits a high-density passive copper backplane spine. It cross-connects every chip directly with 3,600 Gb/s bidirectional NVLink 6 fabric, turning 72 discrete GPUs into one unified memory pool with near-zero latency. 4. Direct-to-Chip 45°C Warm Water Cooling: Dissipating megawatts of thermal power without a wall of deafening fans requires liquid cooling. The rack runs 45°C warm-water direct-to-die cold plates. Warm water eliminates expensive chillers and slashes data center PUE. 5. Cableless Blind-Mate Trays: Inside older clusters, routing thousands of high-speed copper cables by hand took 90 minutes per tray. Rubin introduces blind-mate backplane connectors with zero internal tray cabling, cutting tray assembly time from 90 minutes down to just 5 minutes. The bottom-line payoff: training the same frontier model with a quarter of the GPU count, and slashing inference token costs by up to 10x. Seventy-two chips behaving as one unified superchip. That is how physical limits in silicon are engineered away. Follow @aiunpacked.explains for daily engineering deep dives into AI supercomputing and frontier hardware architectures. 🚀 #nvidia #gpus #verarubin #blackwell #supercomputer #semiconductor #hardware #aiinfrastructure #systemdesign #engineering #tech #datascience

Comments

Comments from the original post on Instagram

No comments yet