March 20, 2024: Nvidia, the US chipmaker with a market value of over $2 trillion, has prolonged its lead in the unreal intelligence sector with the announcement of several groundbreaking technologies at its annual developer conference on Monday.
The company unveiled the “Blackwell” series of AI chips, designed to power the expensive datacentres that train cutting-edge AI models corresponding to the most recent generations of GPT, Claude, and Gemini.
View this post on Instagram
The Blackwell B200 chip is a major upgrade over Nvidia’s previous H100 AI chip, reducing the variety of chips and power required to coach a large AI model like GPT-4 by 75%.
The company also introduced the GB200 “superchip,” which mixes two B200 chips and Nvidia’s Grace CPU on a single board, offering “30x the performance” for server farms that run chatbots like Claude or ChatGPT, while reducing energy consumption by as much as 25 times.
LLM inference and energy efficiency
Nvidia CEO Jensen Huang showcased even larger configurations, corresponding to the GB200 NVL72, a single server rack with 72 B200 chips connected by nearly two miles of cabling, and the DGX Superpod, which mixes eight of those racks right into a shipping-container-sized AI datacentre in a box.
In addition to the brand new AI chips, Nvidia announced Project GR00T, a foundation model for controlling humanoid robots.
This model, paired with the Jetson Thor system-on-a-chip, goals to create autonomous machines that might be instructed using natural language to perform general tasks, even those they haven’t been specifically trained for.
“Project GR00T will help robots understand natural language and emulate movements by observing human actions – quickly learning coordination, dexterity, and other skills with the intention to navigate, adapt, and interact with the true world,” Huang explained.
Nvidia also entered the quantum computing sector with the introduction of a cloud service that uses its AI chips to simulate quantum computers.
While the service is not going to initially be connected to an actual quantum computer, it would allow researchers to check their ideas without the expense of accessing rare and expensive quantum hardware.
In the longer term, Nvidia plans to supply access to third-party quantum computers through the platform.
As Nvidia continues to push the boundaries of artificial intelligence and robotics, the corporate’s latest innovations showcase its commitment to leading the industry and shaping the longer term of those transformative technologies.
