Adapted by:
Javier Pastor
Senior WriterComputer scientist turned tech journalist. I've written about almost everything related to technology, but I specialize in hardware, operating systems and cryptocurrencies. I like writing about tech so much that I do it both for Xataka and Incognitosis, my personal blog.
Karen Alfaro
WriterCommunications professional with a decade of experience as a copywriter, proofreader, and editor. As a travel and science journalist, I've collaborated with several print and digital outlets around the world. I'm passionate about culture, music, food, history, and innovative technologies.
When Nvidia unveiled its new B200 AI chips with Blackwell architecture, the company introduced an AI accelerator called the GB200. Using 36 accelerators, Nvidia created its AI server: the DGX GB200 NVL72. This server has some spectacular surprises in store.
Each node is a beast. Each GB200 accelerator features an Nvidia Grace CPU with 72 ARM Neoverse V2 cores and two B200 GPUs. Combining their power yields a GPU with 1.44 exaFLOPS of FP4 precision.
A cabinet that weighs a ton. The DGX GB200 NVL72 resembles a small, narrow cabinet but is highly dense. This rack weighs 1.36 tons. Inside are 18 Bianca computing nodes in 1U format, each with two GB200s—or four B200 GPUs (18 x 4 = 72). This AI server is estimated to cost about $3 million.
Liquid cooling is essential. These components generate significant heat, making liquid cooling the best option. This system cools the Grace CPUs, B200 GPUs and NVLink chips in the switches, which can run extremely hot due to massive data transfers between accelerators.
Interconnections everywhere. To allow all the GPUs to work together, each of the 36 GB200s uses specialized network cards with fifth-generation NVLink support to connect each computing node to the others. Nine switches provide this enormous number of interconnections.
Almost 2 miles of cable. The system delivers 1.8 TB/s of bidirectional bandwidth between the 72 GPUs in the server. According to The Register, the most surprising aspect is that 1.86 miles of copper cable are inside the cabinet. The switch module alone weighs more than 66 pounds due to the components and the more than 5,000 cables that keep the Nvidia GPUs working in perfect sync.
Why copper? It may seem strange to use copper, especially given this machine’s bandwidth demands. But a fiber optic solution posed challenges: Engineers would have needed electronic components to stabilize and convert optical signals. That would have increased both cost and power consumption.
Can it run Crysis? Each B200 chip is impressive—three times more powerful than the GeForce RTX 5090. The server includes 72 AI-specialized GPUs, showcasing its immense computing power. The B200 also features fourth-generation RT (ray tracing) cores, which could theoretically allow the chips to be used for gaming. But that’s not their purpose. Their gaming performance will likely be as weak as the Nvidia H100.
Sky-high consumption. Although the new chips are far more efficient than the H100s—25 times more efficient, according to Nvidia—this AI server has an estimated TDP of 140 kW. If the average household consumes about 3,000 kWh per year, the server uses the same energy in one hour as a typical household in 17 days. Leaving it on all year would consume as much electricity as 415 average homes in a year.
Image | Nvidia
Related | Nvidia’s Chips Have Long Been a Bargaining Chip in the U.S.-China Trade War—Until Now
View 0 comments