Kathy Gibson reports – The era of generative artificial intelligence (GenAI) is just getting started, with massive innovations and developments still to come.
Carlo Ruiz, vice-president: enterprise solutions and operations at Nvidia EMEA, believes these innovations are going to come in three waves.
The first, which we are already experiencing, is the development of foundational models, or large language models (LLMs).
Ruiz points out that, currently, these models are largely based on growing libraries of human languages. But soon we can expect to see models of machine languages and even the language of the human body or of molecular structure.
The second wave, which has also begun, is the adoption of AI and the emergence of use cases.
“Companies are using AI in-house to create value; and governments are looking to preserve culture using LLMs,” Ruiz points to some of the early applications for GenAI.
One of the concerns as we develop use cases is to ensure it’s done in a smart way, he adds, with issues like privacy, data security, sovereignty and residency to be considered.
The next AI wave, which we can see glimpses of today, is its application in our physical space, in the form of robots, drones, autonomous cars and more.
“This wave has now come of age, with three computing problems having been resolved.”
The first is the technology in the robot itself, and we now have capable systems that can power a small robot or drone. “The compute performance needs to be the same as you had in a data centre a couple of years ago,” says Ruiz.
The other challenge that we have successfully overcome is the AI to provide automation and learning.
The missing link in the middle was the training of these robots, Ruiz says. “But we found that the digital twin, or omniverse environment, allows robots to be trained and to learn in a productive way.
“With these three things coming together, we are ready to deliver performance at scale.”
Nvidia is well-known as the chip provider for the AI world, but Ruiz points out that the company provides an accelerated computing platform that includes the full stack of silicon, system and software.
The foundation of the stack is the silicon, the central processing units (CPUs), graphics processing units (GPUs) and data processing units (DPUs) that drive the full ecosystem, from the data centre to robotic systems, from cloud to edge.
Nvidia adds acceleration libraries and toolkits that allow users to create their own models, and the platforms – Nvidia HPC, Nvidia AI and Nvidia Omniverse – with an AI application framework.
“This is not a theoretical exercise,” Ruiz says. “Often we invested a product because we wanted to use it ourselves.”
By solving problems within its own environment, Nvidia was able to design for predictable performance at scale, with operations and infrastructure manageability and support. It has also developed blueprints for AI workflow management and data science productivity.
And the systems have proven fit for purpose. Ruiz points out that, since the MLPerf benchmark was first set up, Nvidia has dominated the results. “”Nvidia combines hardware and software to ensure continuous performance gains.”
Most Nvidia systems today are built on Nvidia’s Hopper chips, but the introduction of Blackwell is set to dramatically improve performance.
“Blackwell is another step change,” Ruiz says. “People are used to small increments in CPU performance, but this gives incredible leaps in every generation.”
In fact, in the last eight years Nvidia has delivered an impressive 1 000-times performance increase.
To this end, the basic unit of compute is no longer the chip, but the data centre or AI factory, Ruiz says. “You have data coming into the AI factory and have intelligence coming out.
“It is no longer about building a GPU, but about building the entire factory.”
But as processors and data centres gain in performance, they also use more power. Which is why Nvidia takes initiatives to create more sustainable data centres very seriously, Ruiz explains.
“Yes, Nvidia GPU servers consume more electricity – but they also perform at such a high rate that they do traditional compute functions much more efficiently, at lower cost and with better energy efficiency.”
In Europe, Nvidia leads in the Green500 ranking of the world’s most energy-efficient supercomputers, with six out of the top 10.
At the same time, AI can be applied to helping us resolve sustainability issues, Ruiz adds.
Applications include optimising turbines, detecting wildfires and more. “But at Nvidia we thought about what the largest challenges there are, including climate change, and decided our technology needs to make a contribution.”
To this end, the company has created a digital twin of the earth on one of the world’s largest supercomputers – EOS, which ranks number 11 in the world. This digital twin simulates the world and all the elements that are relevant for climate change, collaborating with leading research organisations to predict climate events.