Quantum-tunneling memory could boost AI energy efficiency by 100x

Boffins get excited about building better machine synapses

There's a potential solution on the cards to the energy expenditure problems plaguing AI training, and it sounds simple: just strengthen the "synapses" that move electrons through a memory array. 

Electrical and Systems Engineering Professor Shantanu Chakrabartty and two of his colleagues at Washington University in St Louis, USA, have authored a Nature-published paper explaining how they have used the natural properties of electrons to reduce the energy used to train machine learning models. 

The project had the researchers trying to build a learning-in-memory synaptic array that had digital synapses that operated dynamically instead of statically, such that they only need energy when changing a state, but not to maintain one.

To test their concept, the team built CMOS circuits with energy barriers they said were strong enough to be non-volatile, and which would become stronger (i.e., able to better maintain non-volatility) as the array's training progresses. 

The result, Chakrabartty said, is a more efficient array that could reduce the energy requirements of ML training by 100x – "and this is a pessimistic projection," Chakrabartty told The Register

That 100x improvement is for a small-scale system, Chakrabartty said. Larger-scale models would show an even greater improvement, especially if the memory were integrated with the processor on a single wafer – which Chakrabartty said he and his team are currently working to achieve. 

How to get your digital synapses firing

Machine learning model training is incredibly energy inefficient. Washington University in St Louis said that training a single top-of-the-line AI was responsible for more than 625,000 pounds (283.5 metric tonnes) of CO2 emissions in 2019 – nearly five times what the average car will emit over its life. Other figures say that training a GPT-3 model can burn the amount of energy needed to drive a car to the Moon and back. 

The problem, as the paper sees it, is with the bridges between computing nodes in memory arrays, which the paper likens to the synapses that bridge neurons. In an animal brain, learning strengthens synapses so that they fire more efficiently, but in computing each synapse acts statically. What this means in practical terms is that each time an electron moves through a synapse a "switch" has to be flipped, which spends energy to polarize the synapse, and then has to maintain that energy expenditure to maintain polarity.

The model developed by Chakrabartty and his team gets around that – and creates their more efficient synapse – by using Fowler-Nordheim Dynamic Analog Memory (FN-DAM).

The "FN" portion of FN-DAM refers to the formula that allows an electron to tunnel through a triangular electrical barrier that's electrically isolated (in this instance) by silicon-dioxide barriers. 

Those barriers are strong enough that, even with power removed, the electrons still can't escape. Resupply power in a way that makes that barrier change states, and the electrons trapped in the synapse tunnel away on their journey. 

Chakrabartty said his team's research paper proves that their design is capable, but he warned that FN-DAM still faces a number of barriers to scaling, such as its resolution and measurement precision. ®

Broader topics

Other stories you might like

  • Colocation consolidation: Analysts look at what's driving the feeding frenzy
    Sometimes a half-sized shipping container at the base of a cell tower is all you need

    Analysis Colocation facilities aren't just a place to drop a couple of servers anymore. Many are quickly becoming full-fledged infrastructure-as-a-service providers as they embrace new consumption-based models and place a stronger emphasis on networking and edge connectivity.

    But supporting the growing menagerie of value-added services takes a substantial footprint and an even larger customer base, a dynamic that's driven a wave of consolidation throughout the industry, analysts from Forrester Research and Gartner told The Register.

    "You can only provide those value-added services if you're big enough," Forrester research director Glenn O'Donnell said.

    Continue reading
  • D-Wave deploys first US-based Advantage quantum system
    For those that want to keep their data in the homeland

    Quantum computing outfit D-Wave Systems has announced availability of an Advantage quantum computer accessible via the cloud but physically located in the US, a key move for selling quantum services to American customers.

    D-Wave reported that the newly deployed system is the first of its Advantage line of quantum computers available via its Leap quantum cloud service that is physically located in the US, rather than operating out of D-Wave’s facilities in British Columbia.

    The new system is based at the University of Southern California, as part of the USC-Lockheed Martin Quantum Computing Center hosted at USC’s Information Sciences Institute, a factor that may encourage US organizations interested in evaluating quantum computing that are likely to want the assurance of accessing facilities based in the same country.

    Continue reading
  • Bosses using AI to hire candidates risk discriminating against disabled applicants
    US publishes technical guide to help organizations avoid violating Americans with Disabilities Act

    The Biden administration and Department of Justice have warned employers using AI software for recruitment purposes to take extra steps to support disabled job applicants or they risk violating the Americans with Disabilities Act (ADA).

    Under the ADA, employers must provide adequate accommodations to all qualified disabled job seekers so they can fairly take part in the application process. But the increasing rollout of machine learning algorithms by companies in their hiring processes opens new possibilities that can disadvantage candidates with disabilities. 

    The Equal Employment Opportunity Commission (EEOC) and the DoJ published a new document this week, providing technical guidance to ensure companies don't violate ADA when using AI technology for recruitment purposes.

    Continue reading
  • How ICE became a $2.8b domestic surveillance agency
    Your US tax dollars at work

    The US Immigration and Customs Enforcement (ICE) agency has spent about $2.8 billion over the past 14 years on a massive surveillance "dragnet" that uses big data and facial-recognition technology to secretly spy on most Americans, according to a report from Georgetown Law's Center on Privacy and Technology.

    The research took two years and included "hundreds" of Freedom of Information Act requests, along with reviews of ICE's contracting and procurement records. It details how ICE surveillance spending jumped from about $71 million annually in 2008 to about $388 million per year as of 2021. The network it has purchased with this $2.8 billion means that "ICE now operates as a domestic surveillance agency" and its methods cross "legal and ethical lines," the report concludes.

    ICE did not respond to The Register's request for comment.

    Continue reading
  • Fully automated AI networks less than 5 years away, reckons Juniper CEO
    You robot kids, get off my LAN

    AI will completely automate the network within five years, Juniper CEO Rami Rahim boasted during the company’s Global Summit this week.

    “I truly believe that just as there is this need today for a self-driving automobile, the future is around a self-driving network where humans literally have to do nothing,” he said. “It's probably weird for people to hear the CEO of a networking company say that… but that's exactly what we should be wishing for.”

    Rahim believes AI-driven automation is the latest phase in computer networking’s evolution, which began with the rise of TCP/IP and the internet, was accelerated by faster and more efficient silicon, and then made manageable by advances in software.

    Continue reading

Biting the hand that feeds IT © 1998–2022