Nvidia GB300: presentation of the flagship AI accelerator

Nvidia has revealed all the specifications of its AI accelerator, the Blackwell Ultra GB300. This chip has already gone into mass production, so the first deliveries are expected fairly soon.

Nvidia holds a leading position in the market for flagship AI solutions. For quite some time, the GB200 was the top option on the market, but the new product will soon replace it due to its memory specifications, number of computing units, and many other features.  

Blackwell Ultra Characteristics

The GB300 is equipped with 160 streaming multiprocessors. The TSMC 4NP process technology and the Blackwell architecture as a whole can be considered an improvement over 4N.

According to some reports, the chip will be equipped with fifth-generation tensor cores and will support formats such as NVFP4, FP6, and FP8. In addition, each core will have increased tensor memory capacity. Thus, a twofold increase in performance is expected for certain tasks.

The memory subsystem will have eight 12-slot HBM3E stacks, which will directly increase the capabilities for working with large AI models. 

The platform will be equipped with a PCIe 6 interface, thus providing increased bandwidth. However, energy consumption will also increase: in peak modes, it is possible to use 1400 W.