News Stay informed about the latest enterprise technology news and product updates.

Why 2017 is setting up to be the year of GPU chips in deep learning

GPU technology, once primarily the domain of computer gamers, is getting new legs in systems that run deep learning and AI applications. Experts predict it will be a big part of analytics in 2017.

It's been a while since anyone in the tech world really cared about hardware, but that could change in 2017, thanks...

to the ascendance of deep learning.

According to Forrester analyst Mike Gualtieri, we may be entering a golden age of hardware driven largely by graphics processing units, or GPUs. This alternative to traditional CPU technology is optimized to process tasks in parallel, rather than sequentially. This makes GPU chips a good fit for training deep learning models, which involves crunching enormous volumes of data again and again until the models can recognize images, parse natural speech or recommend products to online shoppers.

"There will be a hardware renaissance because of the compute needs to train these models," Gualtieri said.

Need grows for GPU chips

GPU technology has been around for decades, but only recently has it gained traction among enterprises. It was traditionally used to enhance computer graphics, as the name suggests. But as deep learning and artificial intelligence have grown in prominence, the need for fast, parallel computation to train models has increased.

"A couple years ago, we wouldn't be looking at special hardware for this," said Adrian Bowles, founder of analyst firm STORM Insights Inc. in Boston. "But with [deep learning], you have a lot of parallel activities going on, and GPU-based tools are going to give you more cores."

He said he expects the market for GPU chips to heat up in the year ahead. NVIDIA, one of the first companies to start marketing GPU chips for analytics, teamed up with IBM and Microsoft this year to put its GPUs in server technology aimed at cognitive applications. The biggest player in the processing market, Intel, may also look to get deeper into GPU technology.

GPUs mean changes for developers

Bowles said the growing popularity of this hardware, in some cases, may mean developers will have to learn new ways of working. Since the manner in which a GPU processes data is so different from a CPU, applications will need to be built differently to take full advantage of the benefits. This means developers will need more training to keep their skills sharp going forward.

Similarly, more enterprises will start to build their data architectures around GPU technology, said Tom Davenport, an IT management professor at Babson College in Wellesley, Mass., and co-founder of the International Institute for Analytics, based in Portland, Ore.

In a webcast, Davenport said more businesses are looking at ways to implement image-recognition technology, which is built around deep learning models. Companies like Google, Uber and Tesla are also pushing forward with autonomous vehicles, which lean heavily on deep learning algorithms. The growing role for these types of tasks will increase the demand for GPUs, he thinks.

"We're going to start seeing more computing architectures that focus specifically on GPUs," Davenport said. "That's been a part of the success of these deep learning models for image recognition and other applications."

GPU technology could also enhance the trend of the citizen data scientist, said Paul Pilotte, technical marketing manager at software vendor MathWorks Inc. in Natick, Mass.

He pointed out that Amazon Web Services this year introduced GPU chips to its Elastic Compute Cloud platform. The fact that it's all hosted in the cloud means users don't need to be hardware experts. This lowers the bar to entry and makes the technology available to a wider user base. "The workflows for deep learning, prescriptive analytics and big data will become more accessible," Pilotte said.

Next Steps

GPUs aren't just for gaming anymore

GPUs for deep learning will alter data infrastructure

Big data is boon to deep learning

Deep learning is key to understanding unstructured data

Dig Deeper on Advanced analytics software

Join the conversation


Send me notifications when other members comment.

Please create a username to comment.

How do you plan to implement GPU tools in the year ahead?
In the past data access and network speeds were so much slower than CPU speeds that in big shops CPU parallelism existed, but the focus was on parallel IO and bandwidth.   As Flash memory, PCM, 3D Xpoint and RAM come down in price, IO speed is closer to CPU speed (although still much slower). 

Is this a factor in the interest in more CPU parallelism?
Currently IBM (and I assume Oracle) design their DBMS to make use of parallelism when the SQL is properly written ... and the database is properly designed.  

Will these DBMS be designed to make use of the GPU?  Will effective GPU use depend on good quality SQL (or its alternatives) just as effective use of CPU does?
Interesting! From Microsoft's Direct X framework for graphics, NVidia's PhysX framework for physics processing, to deep learning algorithms, the GPU has come a long way. I wonder what framework will set the foundation for AI programming?
Hmmm  I'm not sure what significance of GPU programming's relationship with deep learning, since most deep learning aka power-series coefficient generation has been achieved for decades without GPUs.  SMIDs provide performance advantages for a very narrow range of computational problems solved with a specific type of parallel solution, where one calculation requires execution over fixed address spaces.  GPUs advantages are quickly diminish over a wider range of algorithms where branches and other types of execution pruning is required.

Amazon has been offering GPUs in an elastic configuration since 2013. See: . Whether the GPU is in chip form or as standalone instances really doesn't make too much of a difference from a development perspective.  My personal preference is the keep the GPU outside the general purpose computational space.  I have 8 Nvidia and 4 Intel Phis (matrix processors) rack mounted connected through PCI-e expansion connections. It provides a more formalized approach towards workflow control.  I usually use the amazon GPUS during development, it keeps the 10KW space heater and jet engine volumes out of my office. 

Until in-execution pruning and zero-impact branches become a true priority for Nvidia, AMD and Intel, GPU actuated deep learning can only provide limited solutions, where even a smaller subset can be considered for deployment in production environments.