Nvidia unveils flagship AI chip, the B200, at GTC 2024

  • NVIDIA introduced the Blackwell GPU at GTC 2024, boasting superior AI training capabilities and energy efficiency compared to its predecessor, Hopper.
  • The Blackwell B200 features over 208 billion transistors and delivers up to 20 petaFLOPS of compute power, while the GB200, comprising two B200 chips, outperforms the H100 by 7x.
  • TSMC’s cuLitho technology, developed in collaboration with NVIDIA and NovuMind Technologies, will be integrated into production, further enhancing semiconductor miniaturization.

At NVIDIA‘s GPU Technology Conference GTC 2024, the company unveiled the next-generation Blackwell GPU, boasting better AI training performance and energy efficiency than its popular predecessor, Hopper. Blackwell comes in two main variants: the B200 and the GB200, which combines two B200 chips with a Grace CPU.

Also read: BDx Indonesia proudly joins NVIDIA’s AI cloud initiative

Also read: Chipmaker Groq and a former AMP VP accuse Nvidia of unfair practises

B200 outperforms the H100 by 7x

The Blackwell B200 features over 208 billion transistors, more than twice the previous chip’s 80 billion transistors, offering improved productivity with nearly simultaneous access to all transistors, delivering up to 20 petaFLOPS of FP4 compute power. NVIDIA claims the GB200, composed of two B200 chips, outperforms the H100 by 7x and accelerates training speed by 4x in benchmark tests based on the GPT-3 model with 175 billion parameters.

NVIDIA states the new GPUs will power AI developments for major players like OpenAI, Microsoft, Google, and other key participants in the AI race.

The Blackwell B200 will utilise TSMC’s 4NP process technology, an enhanced version of the 4N process used in existing Hopper H100 and Ada Lovelace architecture GPUs.

At GTC 2024, TSMC announced the application of the breakthrough chip lithography technology cuLitho, developed in collaboration with NVIDIA and NovuMind, in this year’s manufacturing.

TSMC is partnering with NVIDIA to integrate GPU-accelerated computing

TSMC CEO C.C. Wei stated, ‘We are partnering with NVIDIA to integrate GPU accelerated computing into TSMC workflows, achieving significant performance leaps, throughput enhancements, cycle time reductions, and power consumption reductions. We are transitioning cuLitho into TSMC production, leveraging this computational lithography technology to drive critical components of semiconductor miniaturisation.’

Kinngai Chan, Managing Director of Semiconductor Business at Summit Insights Group, noted that the launch of Blackwell indicates NVIDIA will allocate more capital expenditure than the previous generation Hopper GPU. It is expected that NVIDIA will not only maintain its leadership position but also widen the gap with competitors in the AI field.


Chloe Chen

Chloe Chen is a junior writer at BTW Media. She graduated from the London School of Economics and Political Science (LSE) and had various working experiences in the finance and fintech industry. Send tips to c.chen@btw.media.

Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *