Less than a decade ago, neuroscientist Amir Khosrowshahi was drilling holes and sticking needles into skulls to learn about the human brain.
He turned his knowledge of neuroscience into entrepreneurial success, cofounding Nervana, a startup that helped large companies run neural networks, the technology now driving explosive results in AI. Intel, the world’s largest maker of computer chips, acquired Nervana for more than $350 million just two years later, in 2016.
Since then, Harvard and UC Berkeley-trained Khosrowshahi has emerged as a key AI thinker within Intel, where AI and chips are colliding in new and profound ways.
In an interview with VentureBeat, Khosrowshahi, now CTO of AI, said he is staying at Intel because of a group of researchers there. This team is building a new kind of integrated circuit (IC), filled with transistors that could one day operate at minuscule amounts of energy — as low as 100 millivolts. It would be a step in the direction of the low voltages used for communication in the brain, he said.
The new IC would be a game-changer. It would fuel an explosion of the energy-ravenous AI applications needed to solve some of the world’s biggest problems, including climate change, waste management, and increasingly complex logistics for food and transport systems, Khosrowshahi said.
“What could be more important?” he asked. “This is what’s going to change the field. And in a major way, you know, we’re talking about 30 times performance.”
Most researchers involved in the project — called MESO (magnetoelectric spin-orbit) — acknowledge it could take a decade to make the decisive breakthroughs necessary to bring it to market.
A long-term project with a near-term impact
Khosrowshahi didn’t dispute that it could take a decade to complete the project, but he is also excited about spin-off results MESO is likely to produce within the next two to five years. These could help push AI forward even faster. “We’re going to see a wonderful array of new AI products,” he predicted.
Khosrowshahi’s role in the project is worth clarifying. In many ways, he’s an outsider. He is part of the AI product design team focused on development and projects yielding immediate impact. Moreover, he lacks a background in the physics and circuitry that underpins key research around MESO. While his accomplishments at Nervana give him credibility within the AI field, he’s fighting to catch up with others at Intel who are, by his own admission, way ahead of him. Here he relies on people like Intel senior fellow Ian Young, a circuit designer and lead researcher.
The crisis of CMOS — and the revolutionary alternative
Young, an Australian, has worked for nine years on what he calls “Beyond CMOS” research. CMOS has been the dominant approach to building transistors for the last 35 years, but it is having trouble scaling. While CMOS may allow transistors using “nodes” as small as 7 or even 5 nanometers (a human hair is 100,000 nanometers wide), its power requirements make it expensive — and commercially unviable for certain types of compute, such as that potentially needed by future AI algorithms. Meanwhile, most other research avenues have failed to yield better alternatives, which is why everyone is running around claiming Moore’s Law is dead.
In December, Young’s team (Intel won’t disclose how large the team is), in collaboration with some physics researchers at UC Berkeley, published a paper in Nature that gave the first public outline of the proposed new MESO device. It’s the first time in years anyone has provided a credible, revolutionary alternative to CMOS — one that will catch Intel back up with the Moore’s Law curve. MESO offers promise for a general purpose processor, Young and Khosrowshahi say, so it has to be done regardless.
The device’s 100 millivolts switching would be a fifth of what advanced CMOS devices can operate at. Its attojoule-level switching energy would be about 30 times lower. Moreover, it would be 5 times more dense, and non-volatile, allowing for extremely low standby power.
The group has provided a research roadmap of what is needed to achieve MESO: mainly, breakthroughs in the area of quantum materials, especially in correlated oxides and topological states of matter. To achieve this, Intel realized it needed to open-source the research to allow for collaboration with universities and others. (See our interview with Khosrowshahi and Young, which goes into more details.)
MESO could produce spin-offs, like ASICs for AI
Young is undaunted by the projected time frame. “Why is 10 years a problem?” he asked. Young’s mission is to make sure the MESO device can be used broadly — for all compute purposes. But Khosrowshahi said things will get interesting even before MESO comes to completion, because Intel doesn’t have to wait for the technology to make its way into its general purpose CPU — with its elaborate circuitry, hairy interconnect demands, production planning, and broad applicability. The chip design research will, along the way, include breakthroughs in quantum material structures that can be used for advances in a variety of custom areas. And one of the most promising areas is AI, since — somewhat coincidentally — the MESO transistor’s structure has striking similarities to compute elements in AI neural network architectures.
Indeed, this convergence of MESO and AI is where, during our interview, Khosrowshahi and Young got really excited.
Here’s why: Since MESO’s transistor and interconnects move voltage around in ways that can help data move more efficiently in AI applications, the team can try out some of MESO’s approaches in AI first. Khosrowshahi is convinced this will produce significant improvements in the design of application-specific integrated circuits (ASICs) for AI, and possibly even new products in the datacenter for learning and inference.
AI is simpler in key ways than the CPU, Khosrowshahi explained. AI data flow is significantly easier, for example. “So AI is a great place to experiment on the substrate platform. We’re going to replace the existing substrate for doing matrix multiplication, perhaps with silicon photonics (we will use light). Or we’re going to use room temperature quantum materials. Or we’re going to use analog circuits. There are all sorts of things we can experiment with.”
MESO transistor’s properties reveal the commonalities with AI
To understand the uncanny tie between MESO and AI, remember that computers use strings of digits — 0s and 1s — to store data. A computer’s processor contains billions of transistors that switch off or on to reflect 0s and 1s during computer activity.
The beauty of the MESO project is that it uses quantum materials to do this switching at the nanoscale. MESO uses switches made of ferroelectric material that also have magnetic properties. The magnetic switch solves one of the problems with CMOS, in which switches use much more power.
(Things get even geekier here. For insiders: The switches in the MESO system share similarities with the architecture of neural networks in AI, as Young explains it. With the MESO magnet, multiple inputs can be brought in through a “majority gate,” or thresholding gate. This is analogous to how neural networks use weights to represent the influence of nodes. Similarly, you can see how low voltage and efficient transistor interconnects within MESO can help with AI. The challenge in AI is the massive amount of data movement, and keeping down the energy consumption involved in moving all that data.)
Young has ideas for still more spin-offs, including AI algorithms (not yet conceived) that would otherwise be limited by the amount of available power.
The connection between the need for low energy in chips and the human brain’s own low-voltage state is why Young, the circuit specialist, and Khosrowshahi, the neuroscientist who studied brains, are a good match, the two agree. It was “a big, big interesting coincidence” that they got together, Young said, and their collaboration has profoundly influenced Young’s direction. For Khosrowshahi, the human brain is an “existence proof that better materials exists and that we should keep searching for them.”
(Despite the analogy between MESO and the human brain, and its operations in millivolts, Young said the human brain would still be many orders of magnitude more energy-efficient than MESO.)
The potential reward calls for more investment
Young said the large amount of work required to get to MESO’s end-state demands serious investment.
The industry needs to come together and cooperate more aggressively on this effort, he said. Intel is a member of the Semiconductor Research Corporation (SRC), which is also putting money into the “Beyond CMOS” area. The group includes companies such as IBM, Micron, Samsung, TSMC, and ARM. And President Trump’s recent initiative to support AI includes $2 billion for DARPA, an organization that funds Beyond CMOS research. But more can be done, Young maintained. The research’s benefits to society will be enormous, yet the project remains relatively unknown — even within Intel.
Khosrowshahi said he’s had a tough time getting the ear of some of Intel’s leaders, as the company has many competing priorities and MESO can be difficult to explain. “When I go to executives to talk about device physics and quantum materials at all, it’s quite hard for them to see the relevance. It is a hard story to convey,” said Khosrowshahi. “Why is this relevant? It’s the core technology that runs everything. If you don’t have this, you don’t have anything.”
AI’s momentum could help
Khosrowshahi said there’s already such a virtuous cycle of innovation in the AI space that things are happening faster than people ever expected. The workloads required by AI are providing huge incentives to keep innovating, which makes him believe MESO could realize its potential sooner than anticipated. “I think we’ll get there faster than people expect,” he said. “If people don’t pay attention to it, then I don’t care because it’s a competitive advantage for my company. So either way, it’s good.”
[Khosrowshahi spoke at last year’s Transform event. Transform will be held this year in SF on July 10-11]