GPUs, AI robots

NVIDIA this week at the Computex conference previewed Rubin, the next-generation graphical processor unit (GPU) due out in 2026 that will provide the computational horsepower needed to teach robotic devices the laws of physics.
As the successor to the Blackwell series of GPUs that NVIDIA has just started delivering, the Rubin series will add support for 8-Hi HBM4 stacks, followed by Rubin Ultra GPUs that support 12-Hi HBM4 stacks. A new CPU called Vera is also being developed, along with a Vera Rubin board that combines GPU and CPU into a superchip.

Rubin Series GPUs to Train Robots

NVIDIA CEO Jensen Huang told conference attendees that organizations will use Rubin series GPUs to train robots in virtual gyms hosted in the Omniverse platform provided by NVIDIA how to interact with the physical world. “Computers will learn from each other,” says Huang.

In fact, NVIDIA revealed that a dozen robotics companies, including BYD Electronics, Siemens, Teradyne Robotics and Intrinsic, an Alphabet company, have adopted NVIDIA Isaac, a robotics platform through which NVIDIA shares libraries, AI models and software frameworks. It includes NVIDIA Isaac ROS, a collection of modular ROS 2 packages; NVIDIA Isaac Perceptor, a reference workflow for 3D vision; NVIDIA Isaac Manipulator, a reference workflow for robotic arms; NVIDIA Isaac Sim, a reference application for simulating, testing and validating robots in physical environments; and NVIDIA Isaac Lab, a reference application for reinforcement learning.

Over 100 companies are adopting Isaac Sim to simulate, test and validate robotic applications, including Hexagon, Husqvarna Group and MathWorks. Isaac Lab is being adopted by Agility, Boston Dynamics, Figure AI, Fourier Intelligence and Sanctuary AI.

Organizations leveraging Isaac Perceptor for development of advanced perception-based autonomous mobile robots include ArcBest, BYD Electronics, Gideon, idealworks and RGo Robotics. Companies that have adopted Isaac Manipulator include Solomon, Techman Robot, Vention and Yaskawa.

NVIDIA to achieve its goal is now moving to a one-year development cycle for GPUs that it previously rolled out every two years. Daniel Newman, CEO of the Futurum Group, said NVIDIA is accelerating the pace of innovation in a way that promises to now bring AI to life in robotics platforms.

Assuming Rubin-class GPUs are available in 2026, there might be a whole new generation of humanoid robotics platforms performing tasks that could range from simple household chores to delivering vehicles that other robots have manufactured.

Of course, there are multiple types of AI processors under development so not every robotic platform might necessarily be based on a GPU. Generative AI has become a “killer application” for generative AI because GPUs are designed to process data in parallel. However, they also need to be able to process graphics so in theory a processor designed specifically for AI models might be more efficient.

In the meantime, however, GPUs will remain the dominant class of processors for AI models in the immediate future. The challenge and the opportunity now is to either find ways to optimize their utilization or make greater use of other less expensive classes of processors to run, for example, inference engines at the network edge.