News Hub

NVIDIA unveils ‘World’s Most Powerful AI Chip’ in Blackwell ecosystem

Written by Tue 19 Mar 2024

CEO of NVIDIA, Jensen Huang, announced the ‘world’s most powerful chip for artificial intelligence (AI)’ as part of the semiconductor company’s Blackwell chip ecosystem.

NVIDIA said the Blackwell GPU architecture features six technologies for accelerated computing, intended to facilitate breakthroughs in data processing, engineering simulation, electronic design automation, computer-aided drug design, quantum computing and generative AI.

As part of the Blackwell chip lineup, the B200 represents a new generation of NVIDIA AI graphics processors designed to deliver significant performance upgrades over previous generations.

“This is the most advanced GPU in the world in production today. The excitement of Blackwell is really off the charts,” said Huang.

The Verge reported the B200 GPU is capable of delivering up to 20 petaflops of performance and contains 208 billion transistors. 

NVIDIA emphasised the B200’s ability to handle compute-intensive tasks involved in training AI models, such as feeding data into the model to facilitate learning.

NVIDIA Reveals the GB200

NVIDIA also revealed the Grace-Blackwell Superchip (GB200) which combines two Blackwell GPUs with one Grace CPU to deliver high-performance computing capabilities tailored to the demands of modern AI workloads.

The GB200 includes a specialised transformer engine designed to run transformer-based AI models. This technology is crucial for applications like natural language processing (NLP) and is highlighted as one of the core technologies powering generative AI, like ChatGPT.

“Blackwell offers massive performance leaps, and will accelerate our ability to deliver leading-edge models,” said Sam Altman, CEO of ChatGPT maker, OpenAI.

NVIDIA said the GB200 offers up to a 30 times performance boost compared to an equivalent number of NVIDIA H100 Tensor Core GPUs for large language model (LLM) inference tasks. NVIDIA plans to package the GB200 chips into larger setups like the GB200 NVL72 server, which combines 72 Blackwell GPUs and other NVIDIA components. 

The price of the new Blackwell GPUs is undisclosed. However, analyst estimates suggest that NVIDIA’s previous generation Hopper-based H100 costs between £19,672 ($25,000) and £31,476 ($40,000) per chip, with entire systems priced as high as £157,380 ($200,000).

NVIDIA expects major customers like Amazon, Google, Microsoft, OpenAI, and Oracle to utilise the new chip for both their cloud computing services and internal AI applications.

“By bringing the GB200 Grace Blackwell processor to our data centres globally, we are building on our long-standing history of optimising NVIDIA GPUs for our cloud, as we make the promise of AI real for organisations everywhere,” said Satya Nadella, Executive Chairman and CEO at Microsoft.

In February, NVIDIA said it is building a new business unit to design custom chips for cloud computing firms, including advanced artificial intelligence (AI) processors.

NVIDIA currently controls 80% of the high-end AI chip market. This position has increased its stock market value by 40% this year to £1.37 trillion ($1.73 trillion) after it more than tripled in 2023.

Join Data Centre World Frankfurt

22-23 May 2024, Messe Frankfurt

Experience Germany’s largest gathering of data centre professionals and end-users.

Don’t miss your chance to carve out successful strategies and find solutions that future-proof your next generation of data centres.

Written by Tue 19 Mar 2024

Send us a correction Send us a news tip