Nvidia has announced the Blackwell Ultra GB300, a new professional AI accelerator poised to supersede the GB200. The GB300 boasts enhanced core counts, increased memory capacity, faster input/output capabilities, and higher power consumption compared to its predecessor. The company has confirmed that the GB300 is now in mass production and will soon be available to early adopting customers.
The Blackwell Ultra GPU within the GB300 features 160 streaming multiprocessors, a notable increase from the GB200’s 144. This enhancement brings the total CUDA core count to 20,480. Built on the Blackwell architecture and utilizing the 4NP TSMC node, an advanced iteration of the 4N node, the GB300 promises significant performance gains.
The cores of the GB300 are equipped with 5th-generation Tensor Cores, offering support for FP8, FP6, and NVFP4 formats. These enhancements, coupled with increased Tensor memory, are projected to at least double performance in relevant calculations, according to VideoCardz.
Memory capacity has also seen a substantial upgrade. The GB300 incorporates eight 12-Hi HBM3E stacks, resulting in a total of 288 GB of high-bandwidth memory. This doubles the memory available in the GB200, enabling the platform to handle larger AI models and accelerate their deployment.
To accommodate the increased data throughput, the GB300 utilizes the PCIe 6 interface, which doubles the performance of the previous generation PCIe 5 design, reaching speeds of up to 256 GB/s. This performance comes at the cost of increased power consumption, with the GB300 platform potentially drawing up to 1,400W during peak operation.
While the GB300 is entering mass production for global distribution, Nvidia has stated that, due to export restrictions, the GB300 will not be shipped to China. A less powerful GB30 variant may be made available to China in the future; however, this remains uncertain, given ongoing concerns about Nvidia hardware tracking.
