NVIDIA GB300 will be the green giant’s new graphics accelerator to power the new generation of servers for artificial intelligence. Expected to launch in late 2025, the chip features numerous improvements in performance, memory, connectivity, cooling and other areas compared to its predecessors.
The rise of AI has created a intense demand for server GPUs of NVIDIA, which has increased its market capitalization to more than $3 trillion and made it one of the three most valuable companies in the world. Its new ‘Blackwell’ architecture will debut at CES 2025 in the dedicated GeForce RTX 5000 series graphics cards, but will continue in its professional cards and also in developments for data centers like the one in question.
NVIDIA GB300, a ‘monster’
Known internally by the code “Blackwell Ultra”, this solution aims to provide substantial performance improvements over the GB200. One of the most notable improvements is the increase in memory, since it will go from the 192 GB of its predecessor to the 288GB HBM3e RAM. Additionally, the company changed the architecture from eight layers to twelve layers and the computing board now uses the novel form factor, LPCAMM.
Each B300 chip found in the core of the GB300 (alongside the ‘Grace’ CPU) will require 1,400 watts of powerwhich gives an idea of its performance. Each card promises a 50 percent FP4 performance improvement over GB200. The company’s use of FP4, suitable for inference workloads, is one of the main reasons behind the high expectations for the NVIDIA GB300.
Say that the increase in performance will have an additional cost. Each cabinet requires more than 300 supercapacitors, the production cost of which per unit ranges between $20 and $25. The company will also offer optional battery backup units that cost approximately $1,500 for a complete unit. The system’s cooling infrastructure has been completely redesigned, incorporating advanced water cooling plates and improved quick disconnects into the liquid cooling system.
It is unknown if NVIDIA will eventually roll out a socket configuration for the GB300, which could simplify production and maintenance requirements. The decision also has counterparts, since it would increase the demand for energy and cooling.
Another of the chip improvements will come from the connectivityas the network system has been upgraded from ConnectX 7 to ConnectX-8 InfiniBand, with support for NVIDIA network computing acceleration engines and expansion of optical modules from 800 to 1600 gigabits per second (Gb/s ) data throughput.
NVIDIA GB300 (Blackwell Ultra) expected to arrive at the end of 2025once mass production and shipments of the current GB200 peak in mid-2025 and delays in delivery of solutions that have become essential for AI servers are limited.
