Nvidia "Blackwell Ultra" GB300 GPU to include 288GB of HBM3e memory and draw 1,400W

The big picture: As Nvidia begins rolling out its hotly anticipated Blackwell server GPUs, information regarding an upcoming lineup refresh is starting to emerge. Set for launch in late 2025, GB300 features numerous upgrades in memory, connectivity, cooling, and other areas compared to its predecessor.

Supply chain sources have revealed detailed specifications of Nvidia's GB300 data center GPU to UDN (translated by TechPowerUp). Codenamed "Blackwell Ultra," the chip aims to provide substantial performance enhancements over the company's recently introduced GB200.

One of the most notable improvements is a memory increase, with GB300 growing from its predecessor's 192 GB to 288 GB of HBM3e RAM. Furthermore, Nvidia shifted the architecture from eight layers to 12, and the computing board uses LPCAMM memory.

Each B300 chip at the heart of the GB300 will require 1,400W. Moreover, networking has been upgraded from ConnectX 7 to ConnectX 8, and optical modules have expanded from 800G to 1.6T.

Each card promises a 50 percent FP4 performance improvement over GB200. Nvidia's utilization of FP4, which is well-suited to inference workloads, is one of the primary reasons behind the heightened anticipation for GB300.

However, the increased performance comes with an extra cost. Each cabinet requires over 300 supercapacitors, which carry a roughly $20 to $25 production cost per unit. Nvidia will also offer optional battery backup units costing approximately $1,500 for a complete unit.

Prior reports indicated that the company is considering a socket configuration for GB300, which could simplify production and maintenance requirements. Unfortunately, the decision would also increase power and cooling demands.

Mass production and shipments of GB200 are expected to peak in mid-2025 following delays stemming from late-stage design flaws and overheating problems. Despite the setbacks, orders from Nvidia's clients, including Microsoft, are backlogged for a year. The AI boom has created intense demand for Nvidia's server GPUs, increasing its market cap to over $3 trillion and making it one of the world's three most valuable companies.

Blackwell Ultra is expected to arrive in late 2025, but it remains unclear whether this schedule will collide with Nvidia's next generation, Rubin, which might emerge around the same time. Rubin will utilize TSMC's 3nm N3 process node and HBM4 memory.

Nvidia is also preparing to release its next-generation consumer Blackwell GPUs next year with a likely reveal coming during CES in January.