Nvidia Wants To Democratize AI With Trillion-Parameter Computer Intelligence

Nvidia is putting together a new server processor optimized for tasks such as training artificial intelligence algorithms to be released in 2023. This was announced on Monday at the company’s GTC21 event this week.

A custom platform for AI

Dubbed “Grace”, the processor will allow Nvidia to replace traditional x86 processors from Intel and AMD. Paired with cutting-edge DDR5 memory and a faster version of the company’s NVLink bus, the new architecture is expected to offer a memory-to-GPU performance that is 30 times better – from 64 GB/sec to 2,000 GB/sec, according to charts furnished by Nvidia.

Nvidia chief executive officer Jensen Huang shared on stage about his vision to enable the next generation of artificial intelligence that can approach computer-based “general intelligence”. A custom-designed CPU-GPU platform is necessary to achieve this goal, he says.

Analysts interviewed by Fierce Electronics felt that Nvidia is establishing itself as a leader in AI and machine learning, and democratizing AI by enabling more businesses to leverage the capability of AI.

According to a report on The Next Platform, Nvidia estimates that the Grace architecture “will offer 10X the performance on training natural language models, reducing it from one month to three days, and will allow for real-time inference on a 500 billion parameter model on a single node.”

One step closer to the human brain

The new chip is expected to enable AI computing that is vastly more complex than is possible with today's chip designs, writes AI research analyst Karl Freund in a blog post, noting that the goal is to pursue a trillion-parameter computer intelligence.

As a comparison, he notes that one of today’s largest AI models, the Open.ai GPT-3 has around 170 billion parameters for language processing, requiring over one thousand Nvidia GPUs hosted by Microsoft Azure.

(GPT-3 is a transformer-based language model that can generate paragraphs of text virtually indistinguishable from those written by a well-educated person).

While this sounds like a lot, the human brain is generally thought to have around 150 trillion connections.

“If successful, the Nvidia system would be only 100 times slower than the human brain. While this still does not approach general intelligence, it could be an entire order of magnitude larger than currently planned Exascale systems planned by US DOE labs,” writes Freund.

There are already plans to use Grace in next-generation supercomputers. Nvidia announced that the Swiss National Supercomputing Centre (CSCS) and the U.S. Department of Energy’s Los Alamos National Laboratory will be the first two parties that intend to use Grace-powered supercomputers – to be built by HPE.

Image credit: iStockphoto/Maksim Tkachenko