As part of the GTC 2024 conference, the B200 processor was announced – the basis of a new generation of Nvidia computing accelerators. It is based on Blackwell architecture, has 208 billion transistors and, according to Nvidia, is many times more powerful than its predecessor, the Hopper generation.
Nvidia Blackwell and Hopper processors
Physically, the Nvidia B200 processor consists of two separate 4-nm (TSMC 4NP) chips with 104 billion transistors in each. They are integrated using CoWoS-L technology and communicate via an NV-High Bandwidth interface at a speed of 10 TB/s. On the sides there are eight HBM3e stacks with a total capacity of 192 GB. Memory bandwidth is 8 TB/s. The TDP level of accelerators based on it can be equal to 700 or 1000 W.
The “green” ones have not yet disclosed the exact number of computing units in the B200. The use of new generation tensor cores and support for the fifth generation NVLink interface with a speed of 1.8 TB/s are noted, while using NVSwitch switches you can combine up to 576 processors in the system. When compared with its predecessor based on the Hopper architecture, the B200 is many times greater in terms of computing power.
Along with the new processor, Nvidia introduced a new “superchip”. The Grace-Blackwell Superchip platform combines two B200 accelerators and a 72-core Nvidia Grace ARM processor, and its TDP can reach 2700 W. Systems based on Nvidia Blackwell accelerators will become available to customers this year.
Source:
Nvidia