Details of Nvidia Corp.’s next generation artificial intelligence server, the GB300 platform, have been leaked by Chinese media, and it’s expected to deliver some big improvements in terms of memory, performance, connectivity and cooling.

The world’s most valuable chipmaker is expected to announce the GB300 AI server at its annual GTC event, slated to take place in March 2025, but if it was hoping to keep the new platform under wraps until then, it appears to have failed to do so.

The Chinese language media site UDN claims that supply chain sources have revealed detailed specifications about the new server. The GB300, which will be powered by Nvidia’s most advanced Blackwell graphics processing units, will reportedly come with a significant boost in memory, with 288 gigabytes of HBM3e random-access memory, compared to just 192 GB on its predecessor, the GB200 AI server.

In addition, Nvidia has altered the architecture from eight layers to 12, while the computing motherboard now uses a Low Power Compression Attached Memory Module.

The Nvidia B300 chips that sit at the heart of the GB300 will require 1,400 watts of power, and the networking speed has been upgraded with the switch from ConnectX 7 to ConnectX 8, expanding bandwidth from 800 gigabytes per second to 1.6 terabytes per second.

The report also mentions a 50% increase in FP4 performance compared to the GB200. The decision to use FP4, which stands for four bits of floating-point precision per operation, is said to be one of the main reasons people are so excited about the GB300. The reduced precision gained by moving to FP4 translates to faster compute, reduced data movement and lower power consumption, making it better suited to AI inference workloads.

There are other upgrades too, with the use of a new “slot design” set to be introduced in the GB300 server and a new capacitor tray.

UDN says the upgrades mean that the performance of the GB300 server will improve on that of the GB200 “in all aspects”. It adds that the server stands out as Nvidia’s “next market-grabbing weapon”.

Not surprisingly though, these improved specifications come at a considerable cost, and the GB300 is likely to command a mind-blowing price tag. According to UDN, its supply chain sources estimate that the total production cost of a single supercapacitor in the GB300 will come to between $20,000 and $25,000. With the GB300’s NVL72 AI server cabinet requiring over 300 of those supercapacitors, customers such as Amazon Web Services Inc., Microsoft Corp. and Google LLC can expect to pay at least $7.5 million to fill one up.

What isn’t clear is when the GB300 will go into mass production. Its predecessor, the GB200, is still being ramped up and shipments aren’t expected to peak until the middle of next year, following delays that stemmed from late-stage design flaws in the Blackwell GPUs, which reportedly caused overheating issues.

The setbacks mean that Nvidia has an order backlog of around one year. That’s because the AI boom has led to intense demand for computing capacity, and Nvidia’s GPUs are generally seen as the best in the business. As such, Nvidia’s market capitalization has risen to over $3 trillion, making it one of the three most valuable companies in the world.

While the focus has been on enterprises thus far, Nvidia is also set to roll out its first consumer-grade Blackwell GPUs, and insiders say we may well hear more about these during next month’s Consumer Electronics Show.

Image: SiliconANGLE/Meta AI