Nvidia announces Blackwell, the world's most powerful AI chip.

At a developer event on March 18th, Nvidia announced a series of new products aimed at strengthening its position in the artificial intelligence (AI) market. The chipmaker's stock price has increased 54-fold and revenue has more than tripled since ChatGPT launched its global AI race in late 2022. Nvidia's high-end server GPUs play a crucial role in training and deploying large-scale language models. Companies like Microsoft and Meta have spent billions of dollars purchasing these chips.

The GB200 chip includes 2 GPUs, 1 CPU, and 1 circuit board. (Image: Nvidia)

Nvidia has named its new generation of AI chips Blackwell. The first Blackwell chip – the GB200 – will be available starting later this year. Nvidia is offering customers more powerful chips to boost new orders. Currently, customers are still scrambling to get their hands on the H100 Hopper chip.

"Hopper is great, but we need bigger GPUs," Nvidia CEO Jensen Huang said at the event.

Alongside the Blackwell chip, Nvidia also introduced NIM software to facilitate AI deployment. According to Nvidia officials, the company is gradually becoming a platform provider like Apple and Microsoft, rather than just a chip supplier.

“Blackwell isn’t a chip, it’s the name of a platform,” said Huang. Nvidia Vice President Manuvir Das promised that NIM software would help programmers run programs on any Nvidia GPU, old or new, to reach more people.

Blackwell, Hopper's successor.

Every two years, Nvidia updates its GPU architecture, unlocking new performance growth. Many AI models launched last year were trained on the Hopper architecture, which has been around since 2022.

Nvidia says that Blackwell-based chips like the GB200 significantly upgrade AI performance for AI businesses, with 20 petaflops compared to 4 petaflops on the H100. This processing power allows AI businesses to train larger and more complex models.

The Nvidia GB200 NVLink 2 combines 36 CPUs and 72 GPUs into a single rack. (Image: Nvidia)

The Blackwell GPU is large and combines two separately manufactured dies into a single chip. It's also available as a complete server called the GB200 NVLink 2, which incorporates 72 Blackwell GPUs and other Nvidia components designed for training AI models.

Amazon, Google, Microsoft, and Oracle will sell access to GB200 through cloud services. GB200 combines two Blackwell B200 GPUs with a Grace CPU. Nvidia says Amazon Web Services (AWS) will build a server cluster with 20,000 GB200 chips.

The system can deploy a model with 27 trillion parameters, far larger than even the largest models currently available, such as GPT-4 (1.7 trillion parameters). Many AI researchers believe that larger models with more parameters and data could unlock new possibilities.

Nvidia has not yet announced pricing for the new GB200 or systems containing the GB200. Nvidia's Hopper-based H100 is priced between $25,000 and $40,000 per unit, and entire systems are estimated to cost up to $200,000, according to analysts.

(According to CNBC)

Source