Nvidia has revealed the specifications of the Blackwell Ultra GB300, its forthcoming professional AI accelerator. The GB300 is designed as a replacement for the existing GB200 and features enhancements in core count, memory capacity, and input/output speeds, in addition to an increased power draw.
The current landscape of AI hardware development sees numerous companies globally focused on creating advanced solutions for AI training and inference. Nvidia is positioned at the forefront of this sector, offering hardware designed for efficiency and performance in AI processing. The GB200 has been the company’s top-performing product. The GB300 is expected to provide a performance increase for early adopters.
The Blackwell Ultra GPU, which is integrated into the GB300, incorporates 160 streaming multiprocessors. This is an increase from the 144 streaming multiprocessors found in the GB200. The GB300’s design results in a total of 20,480 CUDA cores. The architecture of the GB300 is based on the 4NP TSMC node, representing an advancement over the 4N node used in the GB200 platform.
The cores within the GB300 are equipped with 5th-generation Tensor Cores, offering support for FP8, FP6, and NVFP4 formats, according to VideoCardz. These cores also feature increased Tensor memory. This configuration is expected to increase performance in relevant calculations. The GB300’s capabilities are designed to address complex computational workloads.
Memory specifications for the GB300 include eight 12-Hi HBM3E stacks. This configuration provides a total of 288 gigabytes of high-bandwidth memory. This doubles the memory capacity of the GB200 and allows the platform to manage larger AI models and accelerate the use of their functionalities. The increased memory capacity aims to improve processing capabilities for demanding AI applications.
The GB300 uses the PCIe 6 interface, which facilitates data transmission. This interface offers double the performance of the PCIe 5 design, with speeds of up to 256 GB/s. Power requirements for the GB300 have increased with the new platform demanding up to 1,400W of power during operation. This increase in power consumption is associated with the enhanced performance capabilities.
The GB300 is currently in mass production and is slated to begin shipping to customers in the near future. However, due to existing export restrictions, the GB300 will not be shipped to China or related territories. A modified, less powerful version, the GB30, may be shipped to China. The decision regarding the GB30 remains unclear given China’s existing concerns regarding Nvidia hardware.
Featured image credit