Nvidia Blackwell GPU-Spotlight on Blackwell GPU, Groot at Nvidia techfest

The world’s third most valued company, Nvidia, on 20th march, announced its most powerful Blackwell chip capable of processing trillion-parameter AI models upto 30 times faster and at one-fourth use of power. Huang addressed a concert-size gathering here to announce the launch of Blackwell chips, set to be used by Amazon Web Services, Dell Technologies, Google, Meta, Microsoft, OpenAI, Oracle, Tesla and xAI.

Blackwell will also power sovereign AI clouds including India’s Shakti Cloud offering by the Hiranandani group’s data centre arm, Yotta Data Services, which localises end-to-end data storage requirements of governments and enterprises.

ET curates an explainer on the processing capabilities of the new AI chipset, what it means for training large language models, among other things announced at the Nvidia GPU Technology Conference, San Jose.

What is the Blackwell GPU?

A graphics processing unit (GPU) is, simply, a computer chip that creates graphics and images by performing mathematical calculations at superfast speeds. That is why GPU’s are suited for high compute requirements of training AI models which are both cost-efficient and energy-efficient.

Huang revealed that Blackwell GPUs are built to train world’s largest AI models with trillion parameters in size. OpenAI’s latest model is going to have 1.8 trillion parameters, which is 10X larger than Chatgpt.

He compared Blackwell’s performance with its predecessor GPU called ‘Hopper’ or commonly the H100. With Hopper, it took 8,000 GPUs, 15 megawatts and 90 days to create the GPT-MoE-1.8T model. With the new system, you could use just 2,000 GPUs and use 25% of the power, he said.

The Blackwell chip is made up of 208 billion transistors and will be able to handle AI models and queries more quickly than its predecessors, Huang said.

Built with 208 billion transistors, the Blackwell chip delivers 1.8TB/s bidirectional throughput per GPU, ensuring high-speed communication among up to 576 GPUs for the most complex LLMs.

Nvidia’s manufacturing partner, Taiwan Semiconductor Manufacturing Co., will use its 4NP technique to produce the product, the company said.

After the announcements, the company’s stock remained largely flat on Monday, but it’s up more than 83% so far this year and more than 241% over the last 12 months.

With a $2 trillion market cap, the chipmaker’s valuation has crossed previous leaders including Amazon and Google.

Global tech luminaries including Microsoft CEO, Satya Nadella, Alphabet and Google CEO Sundar Pichai as well as Cognizant CEO Ravi Kumar S hailed their partnerships with NVIDIA.

“Today we’re expanding our partnership with Nvidia, as we build on our commitment to ensure customers have the most comprehensive platforms and tools across the Copilot stack, from silicon to software, to build their own breakthrough AI capability,” Nadella said in a post on X.

Google’s Pichai said the capabilities of the Blackwell GPU shall be brought to Google Cloud customers.

Cognizant’s Kumar posted that Nvidia’s GenAI platform shall help its “Life Sciences clients to drive efficient and faster Drug discovery by analyzing DNA sequences, predict how proteins will change shape in response to a drug molecule and efficacy of new chemical molecules.”

Project Groot: A foundation model for humanoid robots

The makers of humanoid robots can now use Project Groot, consisting of a new computer based on Blackwell, which will allow robots to understand natural language and copy human movements by observing them.

Developing such skills is “one of the most exciting problems to solve in AI today,” Huang said.

The project involved building a general-purpose foundation model for humanoid robots, which would help teach them to perform everyday tasks.

Among other notable announcements were bringing Nvidia’s 3D computing platform Omniverse to Apple’s recently launched VisionPro virtual reality headset. NvidiaA data centres will send images and videos to the device, providing more lifelike experiences to users.

It also expanded its collaboration with Oracle to help deliver sovereign AI global ..

It also expanded its collaboration with Oracle to help deliver sovereign AI globally. Oracle is the only hyperscaler capable of delivering AI and full cloud services locally, anywhere, it said.