NVIDIA's "Blackwell" Series Upgrade to "Blackwell Ultra"

The latest NVIDIA "Blackwell" series, including B100, B200, and GB200 chips, has just been released to OEMs and hyperscalers. However, the company is already looking ahead to its upgraded "Blackwell Ultra" plans, featuring the upcoming GB300 AI server. According to UDN, this next-generation NVIDIA system will be powered by the B300 GPU chip, running at 1400 W and offering a significant 1.5x improvement in FP4 performance per card compared to its B200 predecessor.

One of the key enhancements is the memory configuration, with each GPU now equipped with 288 GB of HBM3e memory, a notable increase from the previous 192 GB in the GB200. The new design also features a 12-layer stack architecture, an improvement from the GB200's 8-layer setup. The system's cooling infrastructure has been revamped, incorporating advanced water cooling plates and enhanced quick disconnects in the liquid cooling system.

Networking capabilities have also been upgraded, with ConnectX 8 network cards replacing the previous ConnectX 7 generation, and optical modules moving from 800G to 1.6T for faster data transmission. In terms of power management and reliability, the GB300 NVL72 cabinet will standardize capacitor tray implementation, with an optional Battery Backup Unit (BBU) system. Each BBU module costs around $300 to manufacture, with a total BBU configuration cost of approximately $1,500 for a complete GB300 system.

The system's supercapacitor requirements are also significant, with each NVL72 rack needing over 300 units priced between $20-25 per unit during production due to their high-power nature. The GB300, featuring Grace CPU and Blackwell Ultra GPU, will also introduce LPCAMM implementation on its computing boards, indicating the potential for LPCAMM memory standard adoption in servers, not just laptops and desktops. Official launch details for LPCAMM memory configurations are yet to be announced.