The B200 "Blackwell" chip has been designed to deliver 20 PFLOPS (FP4) or 10PFLOPS FP8 of AI performance on a single GPU.
Nvidia Chief Executive Jensen Huang unveiled the new chip at the company’s GTC 2024 developer event in San Jose, California, one of a series of announcements that are intended to maintain Nvidia’s dominance of the fast-growing AI industry.
The new flagship chip takes two squares of silicon the size of the company's previous offering and combines them into a single component.
According to Huang this represents a new class of superchip for an holistic architecture, likened to connecting the right and left side of the brain, with no separation and without requiring programming changes.
It provides 192GB HBM3e and 8TBps HBM bandwidth and 1.18TBps NVLINK. This combination means the ‘superchip’ can expand an AI datacentre scale to above 100k GPUs.
Huang said that the device was 25 times as energy efficient compared to earlier generations.
Major customers including Amazon, Google, Microsoft and OpenAI who are all expected to use the firm's new chip in cloud-computing services and for their own AI offerings.
During his presentation Huang also unveiled a host of new software tools, called microservices, that aim to improve system efficiency by making it easier for a business to incorporate an AI model into its work.
In addition to AI software, Nvidia also unveiled software for emulating the physical world with 3-D models and are intended to help with design in sectors such as automotive and aerospace.
Huang unveiled partnerships with design software companies Ansys, Cadence and Synopsys. He also said that Nvidia's software would be able to stream 3-D worlds to Apple's new Vision Pro headset.
Nvidia also introduced a new line of chips that have been designed for vehicles providing new capabilities to run chatbots inside the vehicle. The Chinese electric vehicle makers BYD and Xpeng have both said that they will be using these new chips.
Towards the end of his presentation, Huang also outlined a new series of chips for creating humanoid robots.
