Learning Curve: How Microsoft Is Using GPUs to Help Machines Learn Faster

by Jay White

The brightest minds at Microsoft Research are making products smarter with NVIDIA GPUs.

Microsoft Research has more than 1,000 world-class scientists and engineers working across research areas in labs worldwide. Their job: make significant product contributions and address some of society’s toughest challenges.

Increasingly, Microsoft engineers and researchers are turning to GPUs to help with their tasks.

Microsoft researchers are finding new ways to put our GPUs to work.
Microsoft researchers are finding new ways to put our GPUs to work.

A growing percentage of their work is focused on machine learning projects – one of the hottest fields in computer science.

The goal of machine learning is to let computers make better predictions by capturing the wisdom of historical data. But history shows that promise is far easier to describe than to achieve.

Three trends are driving a resurgence in machine learning. First, data of all kinds is growing exponentially. Second, researchers have made big improvements in the mathematical models used for machine learning. Finally, GPUs have emerged as a critical computational platform for machine learning research.

These drivers are resulting in game-changing improvements in the accuracy of these models. That’s because GPUs allow researchers to train these models with more data – much more data – than was possible before.

Even using GPUs, the process of training these models by digesting mountains of data takes weeks. Replicating this training process using CPUs is possible – in theory. In reality it would take over a year to train a single model. That’s just too long.

Reducing training time is important because the field is evolving fast. Researchers must accelerate through design and training cycles quickly to keep up. GPUs just cost less, too. The hardware is cheaper and sucks up much less power.

Microsoft Research has just deployed a computer system packed with NVIDIA GPUs. This GPU computing infrastructure is allowing its scientists and engineers to drive innovation and discovery in a range of areas.

These include: computer vision and object recognition, speech analysis, data modeling in fields such as environmental sciences, and machine learning optimization.

In short, the tough stuff.

NVIDIA GPU accelerators are ideal for machine learning, according to Heather Warncke, principal research program manager for big data and machine learning at Microsoft Research.

That’s because GPUs are highly parallelized – they’re built to do many jobs at once – making them perfect for machine learning. That makes them ideal for accelerating the training of deep neural networks, she explained. They’re more energy efficient than CPU-based systems, as well. Plus there are lots of tools – like CUDA – that help them get the most out of the GPUs in the system.

Supercomputing 2014 LogoAs part of this work, we’ve had the opportunity to work with Microsoft Research team members across many different disciplines. We’ve helped them with everything from hardware and systems bring-up to code reviews and performance tuning. And in turn, they’ve shared their insights with us.

If you’d like to learn more about how Microsoft Research is tapping into the power of GPUs, you can hear more firsthand this week. Mark Staveley, a Microsoft senior research program manager, will be among the more than 40 representatives speaking at our GPU Technology Theater at the SC 14 supercomputing confab in New Orleans.

Mark is part of a lineup of talks from individuals using GPUs for a variety of machine learning projects. Baidu’s Bryan Catanzaro will talk about using GPU-powered machine learning for image recognition. Another highlight: Dan Ciresian from IDSIA, an early pioneer for using GPUs with deep learning, will speak about how GPUs and neural networks are aiding in the battle against breast cancer.

Don’t miss it.