Nvidia pushes Blackwell GPU as start of a whole new industry

Nvidia pushes Blackwell GPU as start of a whole new industry

Feature articles |
By Nick Flaherty

Cette publication existe aussi en Français

Jensen Huang, CEO of Nvidia, has detailed its latest graphics processor unit (GPU) technology, from chip to data centre and back down to robotics, pitched as a whole new industry.

“There is absolutely something happening. Industry is being transformed. General Purpose computing has run out of steam, we need a new type of computing in order to scale,” said Jensen Huang in his keynote at the Nvidia GTC technology conference today in California. THis is the first GTC for five years following the pandemic.

“It’s a new industry because the software never existed before, it’s a brand new category and the way you produce the software is unlike anything we have done before, generating tokens and producing floating point numbers at a scale never seen before,” he said.

This is built on the Blackwell GPU, the successor to the current Hopper GPU and named after US mathematician and game theory pioneer David Blackwell who died in 2010.

Huang points to five elements that are key, from the chips to the digital twins that will drive AI foundries and new tools for distributing AI models.

The Blackwell GPU is 208bn transistors with two full sized, reticle-limited die in a single package with coherent memory. “There are no memory locality or cache issues, its just one giant chip with coherent memory,” said Huang.

A key innovation is that the chip includes 100% built in self test of all the gates and memory bits for predictive analysis. “We have the ability to detect a weak node early and swap another node in so we put in a RAS reliability engine with built in self test of every singe gate and every bit of memory on the chip and connected to it and this is the first time we are doing this,” he said. This is performed via the scan chain with a dedicated engine.

Two of these devices are packaged on a board with an ARM-based  Grace processor as the core processing unit.

“It goes into two types of systems,” said Huang. “The first is pin compatible with Hopper with the same power and thermals for HGX.”

These are linked with a new 4nm, 72 port switch chip with 50bn transistors that also includes processing to link the GPUs together in racks, synchronise and update and rebroadcast partial products though the network.

This allows 72 Blackwell GPUs to be connected directly with copper interconnect in a single liquid cooled rack called the GB200 NVL72. “This has 130Tbyte/s back of chassis which is more than the aggregate bandwidth of the Internet,” said Huang.

“Our goal is to continuously drive down the cost and energy of the computing so we can scale up to train the next generation AI models,” he said.

The first chip and systems already have 46 customers at launch. “Every data centre hyperscaler is geared up, ODM manufacturers, every sovereign AI, telcos are ramping up with Blackwell, This will be the most successful product launch in our history.”

On top of this Nvidia has also developed a microservice called NIMS with software containers that allows customers to take their trained models with them to use on any system that uses Nvidia’s Cuda technology.

The final element is robotics, and Nvidia has developed a version of its low cost Jetson card that uses the Blackwell GPU. Jetson Thor is particularly aimed at humanoid robot developers and can be used alongside the Omniverse digital twin tool and Isaac Labs to train the robots.

This is the core of the GR 00T (or Groot) stack for humanoid robots.

“Sensor data created in simulation in NIMs and connect the Isaac NIMS to physical sensors and we’ve made Omniverse simpler to access with four APIs for digital twin capabilities and turned it into an AI so we have taught omniverse USD so that it can look for information semantically such as looking for objects, or conditions or scenarios,” he said.


If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles