Nvidia, the chipmaker, has increased its lead in artificial intelligent with the introduction of a “superchip”, of a quantum computing system and of a suite of new tools that will help to develop the ultimate sci fi dream: general-purpose humanoid robots. We examine what Nvidia is doing, and what this could mean.
The company’s annual development conference was held on Monday. The “Blackwell” AI chip series, which is used to power datacentres, costs astronomically high, are used for training AI models like the latest versions of GPT and Claude, as well as Gemini.
The Blackwell B200 is an upgrade to the H100 AI chip that was previously used by the company. Nvidia stated that to train a GPT-4-sized AI model would require 8,000 H100 processors and 15 megawatts, which is enough power for about 30,000 British homes.
The same training run could be done with just 2,000 B200s and 4MW power. This could reduce the amount of electricity used by the AI industry or allow for the use of the same electricity to power larger AI models.
The company also announced a “superchip” GB200 to go along with the B200. The company squeezed two B200 chips onto a single board along with the Grace CPU to create a system that, according to Nvidia, provides “30x more performance” than the Grace CPU for server farms which run chatbots like Claude or ChatGPT, instead of training them. The company also said that the system will reduce energy consumption up to 25-fold.
By reducing the time that the chips are communicating, they can devote more time to crunching numbers, which makes chatbots sing, or at least talk.
Nvidia would be happy to oblige, as it has a value of over $2tn. Take Nvidia’s GB200 NVL72, a single rack of 72 B200 chips connected by two miles worth of cable. Not enough? Why not look at the DGX Superpod, which combines eight of those racks into one, shipping-container-sized AI datacentre in a box. The pricing was not revealed at the event but, it is safe to assume that you cannot afford it if you need to ask. Even the last generation chips cost around $100,000 each.
Project GR00T, named after Marvel’s arboriform Alien, but not directly linked, is a foundation model developed by Nvidia for controlling humanoid robotics. The foundation model is the AI model that can be used to build specific AI use cases. Examples include GPT-4, which generates text, and StableDiffusion, which creates images. These are the most costly parts of the entire sector, but they are also the engine of innovation because they can be “fine tuned” to specific uses cases in the future.
Nvidia’s foundation model will help robots “understand and emulate natural language, as well as mimic movements, by observing the actions of humans – learning quickly coordination, dexterity and other skills to navigate, adapt and interact with real world”.
GR00T is paired with Jetson Thor, a System-on-a Chip designed to be the brain of a robot. The ultimate goal of the project is to create an autonomous machine capable of being instructed by normal human speech in order to perform general tasks.
Quantum cloud computing is one of the few hot sectors in which Nvidia has not yet made a mark. Microsoft and Amazon have already incorporated the technology, which is still at the forefront of research, into their offerings. Now Nvidia has joined the race.
Nvidia’s cloud won’t be directly connected to a computer that can simulate quantum computing. The service is an AI chip-based simulation of a quantum computing system, which allows researchers to test ideas without having to pay for the expensive (rare) real thing. Nvidia said that in the future, the platform will allow access to quantum computers from third parties.