rewrite this title Nvidia’s Blackwell GPUs move robotics into the physical AI era

Editor’s take: As exciting as generative AI may be, there is another AI-powered technology on the near-term horizon that could prove even more impactful: robotics, or as Nvidia’s Jensen Huang and others have begun calling it, physical AI. The concept of physical AI is that many of the same algorithmic principles used to create large language models for text-based interactions can be applied to learning and replicating physical movements in the real world.

Nvidia has offered robotics-focused hardware platforms and software for more than a decade, but its latest release, the Nvidia Jetson AGX Thor platform (now generally available), marks a significant step forward. For the first time, the company’s latest Blackwell GPU compute architecture is being brought into both industrial and consumer robotics, including humanoid systems.

Traditionally, Nvidia’s robotics platforms have lagged behind its most advanced GPUs. The previous robotics platform, Jetson Orin, was released two years ago and based on the older Ampere architecture, which had first appeared more than three years ago. With the new release, however, Nvidia has finally aligned robotics with its most advanced AI compute engine, the Blackwell GPUs. This represents the first time every one of Nvidia’s vertical offerings is leveraging its latest chips.

Because of the generational leap, performance improvements are substantial. Nvidia claims the Jetson T5000 production module delivers a 7.5-times increase in AI computing power and a 3.5-times improvement in energy efficiency compared to Jetson Orin.

More important than the raw performance gains are the new types of applications the platform enables. Nvidia believes the Jetson AGX Thor will make possible a range of humanoid-style robots suited for commercial use and, eventually, consumer markets.

Nvidia is offering several versions of the Jetson Thor platform. A development kit priced at $3,499 includes a Jetson T5000 board and multiple I/O ports. The T5000 board on its own costs $2,999 and features a 14-core Arm Neoverse V3AE CPU, 128 GB of memory, up to 2,070 TOPS for FP4 calculations, and a 130-watt power draw.

Later this fall, Nvidia will release the T4000, priced at $1,999, with a 12-core Arm Neoverse V3AE CPU, 32 GB of memory, 1,200 TOPS of performance, and a 70-watt power draw. Each Thor platform also includes the new Blackwell Multi-Instance GPU (MIG), which allows the GPU to be divided into multiple virtual parts to handle a wide range of robotic activities and sensor inputs quickly and consistently.

Nvidia Jetson Thor

In addition to hardware, Nvidia is expanding its robotics software ecosystem. The Nvidia Jetson software platform has been updated to support these new boards, and the company’s Isaac Groot humanoid robot foundation models and Nvidia Metropolis for Vision AI are among the tools being integrated.

Of course, when it comes to robots-particularly for consumer applications-there are a huge number of issues to think about beyond just the enabling technology.

The social, psychological, and economic impact of humanoid robots entering everyday life is the kind of thing that’s likely going to take years to fully digest. This is, in many ways, science fiction coming to life, raising questions that must be discussed. To put it simply: Rosie the Robot or the Terminator, which will it be?

Rosie the Robot or the Terminator, which will it be?

Beyond the simplistic good-versus-evil framing, serious questions remain about which consumer applications humanoid robots can realistically address. A digital maid like the Jetsons’ Rosie would appeal to some households, while others would find the idea deeply unsettling. If people worry about “robotic” autonomous driving features in cars, how will they react to real robots in their homes? Moreover, given the cost of the Thor Blackwell computing engine, the first generation of consumer robots could cost tens of thousands of dollars, placing them out of reach for most households.

For these reasons, we are likely to see a gradual progression of robotics-powered consumer devices that ease people into the idea of having robots at home. This may also help soften the sticker shock that will inevitably accompany the technology.

What is less clear, however, is what kinds of non-humanoid consumer devices will emerge and how appealing they will be. Rumors of Apple developing a smart display that swivels to face the user via a robotic arm, for example, hardly sound like a mainstream product.

The industrial outlook is very different. Robotics is already widespread in manufacturing, with rapid growth in autonomous guided vehicles (AGVs) for warehouses and other applications. In these contexts, humanoid robots designed to operate in dangerous or complex environments are compelling and will likely be adopted despite their high cost.

We are still in the early days of physical AI, and unexpected developments may reshape the opportunity, particularly in the consumer market. But with the launch of these new platforms, which can essentially be thought of as advanced robotic brains, it’s clear that the technological tools necessary to create advanced robotics applications are quickly coming into place. And, no surprise, Nvidia has managed to position itself very strongly with a combination of advanced hardware platforms and a wide range of robotics platforms and applications that may help build the kind of moat the company created with CUDA in the world of GenAI.

Lots of questions remain, but the time for starting the discussions on robotics applications – particularly humanoid ones – is now.

Bob O’Donnell is the founder and chief analyst of TECHnalysis Research, LLC a technology consulting firm that provides strategic consulting and market research services to the technology industry and professional financial community. You can follow him on X @bobodtech