Pictured: Kevin Deierling, Senior Vice President of Networking at NVIDIA
In just a few years from now, every business will be an AI business. The reason for this is simple: companies that neglect to embrace artificial intelligence, and fail to embed it deep in their processes, will render themselves obsolete. This is the measure of how pervasive AI is becoming in the world of commerce, and a clear indication of how important it is that senior managers not merely take AI on board, but make a priority out of innovating in its use. Failure to do so is an existential threat.
AI and its awesome power was the subject of a recent interview, conducted for the Business Innovation Leaders Forum and featuring Kevin Deierling, Senior Vice President of Networking at NVIDIA, the chip vendor behind a new generation of accelerated AI platforms. He was in conversation with our Podcast Host Julian Patterson.
“If you look at the human brain, it has something like 100 billion neurons in it,” notes Deierling. “We have chips today that are approaching that sort of number – 100 billion transistors on one chip. The power of the brain, however, is all about the synapses that connect the neurons.”
It’s this ability to cross pollinate at scale that gives modern AI its power, and is reflected, believes Deierling, in how NVIDIA operates: “You can see incredible cross pollination in how we are building things. AI is both a part of our technology and also part of the structure of who we are.”
Innovations in AI are already spelling the end of old certainties, argues Deierling: “Take Moore’s Law, which told us the performance of chips would double every two years,” he says. “It just hasn’t kept up. We can’t count any more on the traditional CPU to double in performance every two years.”
Instead, he believes, we’ve seen a million-x performance rise in AI algorithms: “We’re now thinking outside the box that has traditionally defined the computer and talking about the data set as the new unit of computing, because that’s how large these new AI models are.”
Advances in Deep Learning are indicative of the developments that Deierling speaks of. He describes Deep Learning as a subset of AI, involving neural network models that are so large that they simply can’t be trained in a conventional computer. He cites natural language processing as an example, involving as it does some of the largest AI models in the world.
“Ten years ago, if you tried to talk to a computer, you knew it wasn’t actually speaking and understanding,” he explains. “Today, the state of the NLP art is incredibly good. The models are massive – we’re talking the power of hundreds of billions, ultimately trillions of neurons. That requires the kind of Deep Learning where people aren’t writing the programmes anymore. Instead you’re feeding a massive amount of data to a computer, and the computer is writing the programme. That’s what Deep Learning is all about.”
Training at the next level
Training and inference are the twin planks on which modern AI rests, says Deierling, and an approach that unifies the two is needed: “Take the example of the self-driving car,” he says. “You need to train it on massive data sets, not least because the roads are different in the UK than they are in the US. Learning is a massive part of that.”
The execution of the model, he says, then depends on an inferencing server: “At NVIDIA, we have a single architecture that spans across learning and Deep Learning and into the inferencing. You need to do all that instantaneously in a self-driving car.”
He expects this type of all-encompassing, real-time AI to pop up in other use cases, for example medical imaging: “We will see it both at the clinical level at the research level, but also in your doctor’s office where AI will be part of your standard check-up procedure,” he predicts. “Your doctor will look at some data and she’ll know exactly what’s happened because AI is augmenting all of the things that she has studied.”
Deierling cites another cutting edge example of advanced AI in action, as demonstrated at a recent NVIDIA conference: “We recreated a BMW factory as a digital twin,” he explains. “This twin world, or omniverse, follows all the laws of physics, with gravity and light reflections, with robots, and human beings. If your business model is to be replicated throughout the world, you can’t just have one factory. Before you build a new factory, you can use the digital twin to see how real human beings engage with sensors and interact with robots on the floor. You can project yourself through augmented reality into the factory and figure things out before you spend billions of dollars to build it. This piece of equipment is too low, and if I build it this way it’s going to hurt my back.”
He says this kind of 5G-connected omniverse applies equally if you’re building a car or a plane or a building, or within any type of collaborative environment: “Now you can build a virtual world that is accurate down to the lights, the gravity, the physics, and everything is going to be real. Manufacturing is going to be completely revolutionised, with more and more synthetic data used to train your models.”
Fixing the cyber crisis
Deierling argues that AI innovation is not just set to change how we make things and manage information, it will also have a bearing on how we keep essential data and services safe. He describes NVIDIA’s Morpheus platform, designed to change how cybersecurity is approached: “What’s neat about Morpheus is that it’s so adaptive,” he says. “Take a situation where you don’t know what the threat is, let’s call it a zero day threat. It’s brand new, it’s never been detected before, so you can’t look it up in a database of malware. Naturally you want to warn the world as quickly as possible, to tell others not to let it in. But if you’ve never seen it before, how can you follow its behaviour and detect that it’s bad? We’re using AI to do that. AI is also going to help prevent accidents in the workplace. It’s going to be able to look at things in a very intuitive way that you can’t do using a fixed rule set and a fixed algorithm.”
So what of the future? What does innovation in the field of AI have in store for us way down the line? Deierling cites Hal Varian, Chief Economist at Google. Varian, he said, identified a simple way to forecast the future: just look at what rich people have today. And then a decade later you’ll see that become common in the middle class. And then a decade after that, everyone will have it.
The AI innovation that excites Deierling the most is the kind that makes life less complex: “I have a smartwatch, and I have phones and I have different pieces of IoT equipment,” he says. “I scanned my network at home the other day, and I was shocked at how many different things I have. Even my camera is highly electronic and connected. To tell you the truth, I can’t use any of them very well because they’re too complicated. For example I can’t figure out how to change to the maps that I have built into my watch.”
It’s Deierling’s perception that the wealthiest people out there can afford tech that doesn’t force them to become IT managers in their own home or in their own car: “I believe they’re now interacting with their car in a very human way, talking to it, and the car is understanding what’s happening. I predict that the next explosion that we’re going to see in AI is to simplify the complexity that’s already built into the devices we have. We need simplicity to invoke all of these capabilities. I want to talk to my computer, I want it to recognise human language and do the simple things that I want it to do. That’s real progress.”
- Kevin joined NVIDIA when the company acquired Mellanox in 2020. The combined organisation, which made $17 billion in its most recent financial year, makes the processors that power the global computer gaming industry, and now also help to meet the appetite for compute power in the cloud. Kevin has contributed to multiple technology standards, and holds more than 25 patents in areas as diverse as security, wireless communications, video compression and DNA sequencing. He’s an expert on semiconductor design, and he holds a BA in solid state physics from the University of California at Berkeley.
The podcast is available on Blubrry here.
And on iTunes here.
This NVIDIA event should be regarded as unmissable.
And this event should also be diaried by anyone with a stake in AI and networks.
These blogs will provide useful additional info.
By Guy Matthews, Forum Content Editor
Responses