Nvidia unveils super chip using TSMC technology at annual conference

Taipei, March 19 (CNA) Nvidia kicked off its annual GPU technology conference (GTC) in San Jose, California on Monday (local time) by unveiling its latest AI super chip that uses TSMC's 4-nanometer process technology.

During his two-hour keynote speech that was live-streamed, CEO Jensen Huang (黃仁勳) introduced the company's new flagship AI chip 'Blackwell' -- named in honor of mathematician and game theorist David Harold Blackwell -- and said it will speed up AI computing.

The superchip is packed with 208 billion transistors and manufactured using TSMC's N4P (4nm) process technology. As a comparison, the graphics processing unit (GPU) introduced by Nvidia in 2022, Hopper, had 80 billion transistors.

A GPU is an electronic circuit that can perform mathematical calculations at high speed and was initially designed to accelerate image processing.

Calling Blackwell "a very big GPU," Huang said it is increasingly necessary because AI models larger than the one used by ChatGPT will be needed as "we are going to train [AI] with multimodality data -- not just text on the internet but images and graphs ... and videos."

To demonstrate the advantage of the new GPU, Huang explained that to train a GPT model -- like ChatGPT -- it would take 8,000 Hopper GPUs, 15 megawatts, and 90 days; but with Blackwell, with the same length of time, "it would only take 2,000 GPUs and four megawatts of power."

The company at the same event also announced another element to its partnership with TSMC.

Huang said TSMC will use Nvidia's new computational lithography platform in the manufacturing of its new chips.

"Computational lithography is a cornerstone of chip manufacturing," Huang said, adding that Nvidia's partnership with TSMC "applies accelerated computing and generative AI to open new frontiers for semiconductor [production]."

Lithography, or photolithography, is a step in the chip-making process involving using light to transfer a pattern onto a wafer.

TSMC CEO C.C. Wei (魏哲家) was quoted in Nvidia's press release as saying that working with the company to integrate GPU-accelerated computing in the TSMC workflow "has resulted in great leaps in performance, dramatic throughput improvement, shortened cycle time, and reduced power requirements."

Another Taiwan-related announcement made by Huang was related to Nvidia's new generative AI model "CorrDiff," a weather forecaster, which so far has been "optimized for Taiwan."

But soon the technology will be available in different regions of the world, the introduction video played at Nvidia GTC said.

"Current AI forecast models can accurately predict the track of storms, but they are limited to 25-kilometer resolution which can miss important details.

"Nvidia's CorrDiff is a new generative AI model that can be used to have 2-kilometer resolution with 1,000 times the speed and 3,000 times the energy efficiency of conventional weather models," it continued.

Taiwan's Central Weather Administration plans to use the AI model so it can predict where typhoons will hit more accurately.

"Taiwan is a critical component of the global supply chain, and flooding risk analysis and evacuation preparedness are core to our mandate," Cheng Chia-ping (程家平), head of the administration, was quoted as saying in the Nvidia press release.

The annual Nvidia GTC will run until March 21.