Thought Leaders in Artificial Intelligence: Steve Scott, CTO of Cray (Part 2)

Steve Scott: The way people have used Cray and other high-performance supercomputers is, you have a bunch of equations that present a model for the natural world whether that’s equation of airflow across an airplane wing or equations dictating the molecular dynamics involved in drug discovery.

You iteratively solve these equations spread across these points in space. You’re figuring out what goes on by solving these large numbers of equations that represent the real world. In AI, we have a different way of calculating results. Deep neural networks are more of data-driven versus mathematically-driven techniques where you have these layers of of artificial neurons. They’re taking inputs in.

It gets trained to produce some sort of result whether it’s recognizing an object or classifying something based upon the data. Of course, you’re just learning based on the data rather than on these equations. How do these two things come together? There are various ways in which AI can accelerate computation.

For example, you might be predicting the flow through a turbine or a jet engine. Normally, you’d be doing this by iteratively solving the Navier-Stokes equations. The computation proceeds by solving these equations, getting a better of the flow field though this turbine, and then solving them again until it converges to a known state.

What we can do with AI is replace part of that computation by an initial estimate of what the flow would look like just based on the deep neural network. It’s much much faster. It’s approximate. People aren’t necessarily trusting of the results because there’s this black box effect, but you can get an initial approximation of the result very quickly. Then you can use that to precondition your iterative solver.

You’re effectively using AI to come up with an approximate solution. Then your iterative numerical solution goes much, much faster. This preconditioning idea can speed up any kind of iterative algorithm. You can also replace part of a simulation with a deep neural network. In weather forecasting, one of the more expensive components is the radiation transport part of the calculation.

It doesn’t need to be exact and it’s rather expensive to compute and you can replace that part of the computation with the results of a deep neural network and make your weather code run much faster. There are a number of examples where people can use AI to steer a computation, precondition a computation, replace part of a computation, and make it run faster.

Every one of our large customers that does this traditional simulation and modeling is looking at applying AI to that workflow and making it run faster. That’s the approach of using AI to help HPC. The other thing that we’re interested in is in using HPC techniques to help AI.

That’s basically recognizing that many of the problems in AI are very similar to high-performance computing problems that we’ve solved traditionally. As a company that’s just good at doing scalable high-performance computing, we can bring some of those techniques to make straight AI workloads run faster.

This segment is part 2 in the series : Thought Leaders in Artificial Intelligence: Steve Scott, CTO of Cray
1 2 3 4 5