Leaks reveal beefed-up specifications for Nvidia’s next-gen GB300 AI server

Particulars of Nvidia Corp.’s subsequent technology synthetic intelligence server, the GB300 platform, have been leaked by Chinese language media, and it’s anticipated to ship some huge enhancements by way of reminiscence, efficiency, connectivity and cooling.

The world’s most dear chipmaker is anticipated to announce the GB300 AI server at its annual GTC occasion, slated to happen in March 2025, but when it hoped to maintain the brand new platform underneath wraps till then, it seems to have failed to take action.

The Chinese language language media web site UDN claims that provide chain sources have revealed detailed specs in regards to the new server. The GB300, which will probably be powered by Nvidia’s most superior Blackwell graphics processing units, will reportedly include a big enhance in reminiscence, with 288 gigabytes of HBM3e random-access reminiscence, in contrast with simply 192 gigabytes on its predecessor, the GB200 AI server.

As well as, Nvidia has altered the structure from eight layers to 12, whereas the computing motherboard now makes use of a Low Energy Compression Hooked up Reminiscence Module.

The Nvidia B300 chips that sit on the coronary heart of the GB300 would require 1,400 watts of energy, and the networking pace has been upgraded with the change from ConnectX 7 to ConnectX 8, increasing bandwidth from 800 gigabits per second to 1.6 terabits per second.

The report additionally mentions a 50% enhance in FP4 efficiency in comparison with the GB200. The choice to make use of FP4, which stands for 4 bits of floating-point precision per operation, is outwardly one of many primary causes individuals are so excited in regards to the GB300. The diminished precision gained by shifting to FP4 interprets to quicker compute, diminished information motion and decrease energy consumption, making it higher suited to AI inference workloads.

There are different upgrades too, with using a brand new “slot design” set to be launched within the GB300 server and a brand new capacitor tray.

UDN says the upgrades imply that the efficiency of the GB300 server will enhance on that of the GB200 “in all features.” It provides that the server stands out as Nvidia’s “subsequent market-grabbing weapon.”

Not surprisingly, although, these improved specs come at a substantial value, and the GB300 is more likely to command a mind-blowing price ticket. In keeping with UDN, its provide chain sources estimate that the overall manufacturing value of a single supercapacitor within the GB300 will come to between $20,000 and $25,000. With the GB300’s NVL72 AI server cupboard requiring over 300 of these supercapacitors, prospects comparable to Amazon Net Providers Inc., Microsoft Corp. and Google LLC can anticipate to pay a minimum of $7.5 million to fill one up.

What isn’t clear is when the GB300 will go into mass manufacturing. Its predecessor, the GB200, continues to be not at peak manufacturing and shipments aren’t anticipated to peak till the center of subsequent 12 months, following delays that stemmed from late-stage design flaws within the Blackwell GPUs, which reportedly precipitated overheating points.

The setbacks imply that Nvidia has an order backlog of round one 12 months. That’s as a result of the AI growth has led to intense demand for computing capability, and Nvidia’s GPUs are typically seen as the perfect within the enterprise. As such, Nvidia’s market capitalization has risen greater than $3 trillion, making it one of many three most dear corporations on the planet.

Although the main focus has been on enterprises up to now, Nvidia can also be set to roll out its first consumer-grade Blackwell GPUs, and insiders say we could effectively hear extra about these throughout subsequent month’s Client Electronics Present.

Picture: SiliconANGLE/Meta AI