Future AI hardware development may drive materials science R&D.

WhatsApp Group Join Now
Telegram Group Join Now
Instagram Group Join Now

[Adobe Stock]

Despite all its promise, artificial intelligence is still stuck in its teenage phase. Tech has excelled at certain tasks, but lacks maturity.

While creative AI software like GPT-4 has captured the public imagination, the hardware needed to unlock AI's true potential has lagged behind. “I don't think the hardware is there yet,” said Pradyumna (Pradi) Gupta, Ph.D., founder and chief scientist of Infinity Labs, which provides testing services for R&D initiatives. While the latest crop of graphic processing units (GPUs) enable the creation of trillion-parameter-scale deep learning models, Gupta believes the memory component of hardware is the “real bottleneck” in AI.

Breaking down the wall that limits AI's potential.

One name for this phenomenon in computer architecture is the “memory wall”. “The tendency to have more cores on a chip exacerbates the situation, as each core enjoys a relatively narrow channel for shared memory resources,” explained Sally A. McKee and Robert W. Wisniewski. Encyclopedia of Parallel Computing.

Pradyumna (Pradi) Gupta

As AI models become more complex, they require larger amounts of data processing and, consequently, lower latency, higher bandwidth memory. “Obviously, better logic chips are coming out, but memory is still the same old memory,” Gupta said. “There are new ones. [memory] Technologies are coming, and when they come, I think it will be a real revolution.

Growing AI interest could drive demand for next-generation memory. Gartner predicts that demand for high-bandwidth memory will grow eightfold from 2022 to 2027.

Pushing the limits of hardware

“The way you think is that a lot of people, even in the chip industry, are thinking. They think this could be the end, which is surprising,” Gupta said, referring to the observation. The number of transistors on a microchip doubles roughly every eighteen months to two years, which is known as Moore's Law, Hoy said. “Companies like TSMC, which are the most technologically advanced, are making chips that go down to three nanometers. The A14 chip in the iPhone 12 and the A15 chip in the iPhone 13, for example, are from TSMC. Manufactured using 5nm process.

“TSMC has a roadmap to cross and go below one nanometer. Intel also has plans for sub-nanometer technology,” Gupta said. “When they get to that level, they'll be playing with individual atoms.”

Gupta emphasized the importance of this development: “Ten years ago, no engineer in the semiconductor industry, not even the people who make the chips, would have thought that we could get as close to the atomic level as three nanometer technology. Will go now.”

A unique perspective

As a former senior scientist at Corning who helped develop the Gorilla Glass used in the first smartphones, Gupta has a unique perspective on the intersection of materials science and technology. “If you think about glass, when you drop it, it breaks, right?” They said. But Gorilla Glass, a product that has been in development for generations, aims to engineer a glass that can survive even a single drop. “You're making something out of glass that's very different from its basic nature,” Gupta said.

In the world of materials science and deep tech, the R&D process can present a unique set of challenges compared to software development. “In material science, pharma, or any field that involves the physical sciences where you have to make something, it's very different,” Gupta said. “You follow the same loop: you think about something, design an experiment and build a physical model. A lot of things can go wrong. You take something, and the real world gives you something else. It's not a computer program, it's a real thing.”

Democratizing access to testing resources

But even as computing power advances, Gupta argues that, in addition to memory limitations, access to sophisticated testing equipment remains a major barrier to innovation. That's where his company, Infinita Lab, comes in.

“We have a network of about 2,000 labs across the U.S., and that number is growing,” he said. “What we're doing is democratizing access to nearly unlimited content testing services at a very low cost,” Gupta explained. “The problem we solve is that a startup or individual engineer often needs to do some testing, but either doesn't have the equipment at home, doesn't know where to go to get the test done. is, or the place they need is inaccessible, like a university lab that is difficult to access from the outside.

Infinita Lab provides a variety of lab testing services to evaluate materials, products and processes. For example, it provides metrology testing using microscopy and spectroscopy, materials testing to evaluate the properties of various materials such as metals and polymers, and product testing under normal and stressed conditions.

To offer this variety, it partners with labs across the U.S. “You can say. [the model is] Like Amazon or Uber,” he said. That is, the company makes a product or service available, but doesn't necessarily own the inventory.

Shaping the world of the future

Returning to the topic of artificial intelligence, Gupta sees significant potential for the technology in material science. “There are a lot of unknowns in the process that you can potentially drive with AI,” he said. “AI can have a significant impact on every part of the process, from designing experiments to analyzing data and determining whether a particular sample preparation or recipe should be tested. All these questions can be answered by AI. will be delivered in a much more efficient and optimal manner, which will drive R&D at a much faster pace.

WhatsApp Group Join Now
Telegram Group Join Now
Instagram Group Join Now

Leave a Comment