Blackwell Ultra: Nvidia GB300 with 288 GB HBM3e and 1.4 kW Pro GPU

Also in Taiwan we have some technical data regarding the Upgrade with Blackwell Ultra. The Mid-Cycle-Refresh of Nvidia’s Blackwell architecture for Computex in June has already brought details, such as the details of the GB300 like the HBM3e model (288 GB) or the pro GPU (1,400 watts) and the light.

As Nvidia das Tempo bei der Eineer Lösungen für das KI-Supercomputing will be available soon, the manufacturer must address a year to attract investors. During the Computex keynote in June, the war of these years CEO Jensen Huang was guilty and lost his mind for the next few days over the years Blackwell-Nachfolger Blackwell Ultra, Rubin and Rubin Ultra.

Blackwell Ultra with 288 GB HBM3e

I’m going to recommend the mid-cycle refresh of the Blackwell Ultra the month after. For this reason, Nvidia is available on the official product page “8S HBM3e 12H”. This is the case: It includes 8 HBM3e batteries with 12 layer jewels for installation. The professional GPU is 192 GB to 288 GB, but it is available over time. Economic daily aus Taiwan sous Verweis auf die Lieferkette.

The 288 GB price does not exceed expectations, because Blackwell (without Ultra) has an 8 Stacks HBM3e with 8 jewel layers and it reaches 192 GB. 50 percent more layer with Blackwell Ultra for 50 percent storage more memory and approximately 288 GB.

TDP is moderate

In the Bericht der Economic daily This is the first GB300, as well as the Nachfolger of the GB200 Superchips. So bezeichnet Nvidia das Board with two B200 GPUs and a Grace CPU. 36 these cards are represented by the GB200 NVL72, a single package, waterproof rack with 72 Blackwell GPUs, 36 Grace CPUs, 13.5 TB HBM3e and up to 1,440 PetaFLOPS Leistung (FP4 Tensor Core Sparse). This Superchip-Board features a configured TDP of 2700 watts. This is what happens with the click on Blackwell Ultra.

The B300-GPU comes with a TDP of 1,400 Watts. The next time you use it, this is not the case but it is also B200. Two Blackwell-Ultra-GPUs have a maximum power of 2,800 watts, but they now need to use the Grace-CPU and other components on the board, in order to use the GB200 superchip. Dass der Verbrauch pro Board höher sein wird, ist absehbar, der Unterschied wird gemessen an dem Leistungszuwachs aber nicht dramatisch ausfallen. The FP4 module is fair with Blackwell Ultra for using the Sparsity module at 50 percent on 2,160 PetaFLOPS.

LPCAMM for the Grace processor

The absence of HBM3e GPUs includes a GB200 superchip and 480 GB LPDDR5X for the Grace processor. Here we offer the next generation of LPCAMM module, the only solution for laptops, desktops and servers.

Änderungen bei Kühlung et Energiespeicher

For the new AI server with Blackwell Ultra, the solution for the water system is the solution for universal quick connections (UQD), the modular system has been designed. In the new racks are located in the central place for installed battery backup units (BBU) and supercapacitors, with this energy source absorbed.

Networking at 1.6 Tbps throughput

Außerdem soll Nvidia beim Networking die nächste Verdoppelung auf 1.6 Tbit/s gehen. Die Economic daily bring an upgrade from the new ConnectX-7 InfiniBand adapters to the in-game ConnectX-8. The Computex roadmap was launched in June with 1.6 Tbps for the first time for the coming years, the platform -Rubin form and the ConnectX-9 have been implemented. The ConnectX-8 is also compatible with Blackwell Active solutions for simple installation and 800 Gbps bandwidth for use.