IBM is announcing today an enhancement to its PowerAI software platform aimed at facilitating the practical scaling of AI models on today’s fastest GPUs. Scaling to 256 GPUs with its new distributed deep learning (DLL) library, IBM reports that it has bested previous records set by Google and Facebook on two well-known image recognition workloads.
“This is one of the bigger breakthroughs I have seen in a while in all of the deep learning industry announcements over the last six months,” said Patrick Moorhead, president and principal analyst of Moor Insights & Strategy. “The interesting part is that it is from IBM, not one of the web giants like Google, which means it is available to enterprises from on-prem use using OpenPower hardware and PowerAI software or even through cloud provider Nimbix.”
The crux of the announcement is a new communication algorithm developed by IBM Research scientists and encapsulated as a communication library, called PowerAI DDL. The library and APIs are available today as a technical preview to Power users as part of the PowerAI version 4.0 release. Other efforts to improve multi-node communication have tended to focus on only a single deep learning framework, so it’s notable that the PowerAI DDL is being integrated into multiple frameworks. Currently TensorFlow, Caffe and Torch are supported with plans to add Chainer.
Customers who don’t have their own Power systems can access the new PowerAI software via the Nimbix Power Cloud.
“Like the hyperscalers and large enterprises, Nimbix has been working to build distributed capability into deep learning frameworks and it just so happens that what IBM is announcing is effectively a turnkey software solution that implements that in multiple frameworks,” said Nimbix CEO Steve Hebert.
“This is truly an HPC technology,” he continued. “It’s taking some of the best software components of traditional HPC and marrying those up with AI and deep learning to be able to deliver that solution. Our platform is ideally suited for scaling out in the HPC sense, very low latency for codes that get that linear scaling of problem sizes. That means for deep learning we can start to tackle enterprise-class deep learning problems basically on day one. For this to become available to any company or consumer outside of [the big hyperscalers], like Google, Baidu, etc., it really democratizes access to everybody.”
The multi-ring communication algorithm within DDL is described (see IBM Research paper) as providing a good tradeoff between latency and bandwidth, as well as being adaptable to a variety of network configurations. The full method is proprietary but section 4 of the paper provides a fairly detailed description of the library and algorithm.
The current PowerAI DDL implementation is based on Spectrum MPI. “MPI provides many needed facilities, from scheduling processes to basic communication primitives, in a portable, efficient and mature software ecosystem” state the researchers, although they add the “core API can be implemented without MPI if desired.”
To evaluate the performance of its new PowerAI Distributed Deep Learning library, IBM performed two experiments using a cluster of 64 IBM “Minsky” Power8 SL822LC servers, each equipped with four Nvidia Tesla P100 GPUs connected through Nvidia’s high-speed NVLink interconnect. The systems occupied four racks (16 nodes each), connected via InfiniBand.
IBM reports that the combination of its Power hardware and software offers better communication overhead for the Resnet-50 neural network using Caffe than what Facebook recently achieved with the Caffe2 deep learning software. The IBM Research DDL software achieved an efficiency of 95 percent using Caffe on its 256-GPU Minsky cluster whereas Facebook achieved 89 percent scaling efficiency on a 256 NVIDIA P100 GPU accelerated DGX-1 cluster using the Caffe2 framework. Implementation differences that could affect the comparison, e.g., Caffe versus Caffe2, are discussed in the IBM Research paper.
In the second benchmark test, IBM Research reported a new image recognition accuracy of 33.8 percent for a Resnet-101 neural network trained on a very large data set (7.5 million images, part of the ImageNet-22k set). The previous record published by Microsoft in 2014 demonstrated 29.8 percent accuracy.
IBM Research fellow Hillery Hunter observed that a 4 percent increase in accuracy is a big leap forward as typical improvements in the past have been less than 1 percent.
Further, with IBM’s distributed deep learning approach, the ResNet-101 neural network model was trained in just seven hours, compared to the 10 days it took Microsoft took to train the same model. IBM reported a scaling efficiency of 88 percent.
Sumit Gupta, vice president of AI and HPC within IBM’s Cognitive Systems business unit, believes the increased speed and accuracy will be a huge boon to enterprise clients. “Part of challenge has been if it takes 16 days to train an AI model it’s not really practical,” he said. “You only have a few data scientists when you work in a large enterprise and you really need to make them productive so bringing down that 16 days to 7 hours makes data scientists much more productive.”
Certain applications are particularly time-constrained. “In security, military, fraud protection, and autonomous vehicles you often only have minutes or seconds to train a system to deal with a new exploit or problem but currently it generally takes days,” said market analyst Rob Enderle. “This effectively reduces days to hours, and provides a potential road map to get to minutes and even seconds.” It’s scenarios like these that make buying Power Systems to speed deep learning far easier to justify, he added.
The list of use cases seemingly grows longer by the day. Recommendation engines, credit card fraud detection, mortgage analysis, upsell/cross-sell to retail clients, shopping experience analysis are all getting a lot of attention from IBM’s customers.
“The giants like Microsoft and Google and others who have consumer apps, they obviously are getting on the consumer platform a lot of data all the time. So their use cases in many cases are very obvious, finding images of dogs in Google photos,” for example, said Gupta. “But we see enterprise clients have lots of data and lots of use cases they are now getting around to using these methods.”
The next step for IBM researchers is to document scaling beyond 256 GPUs as their current findings indicate that is feasible. “We don’t see a reason why the method would slow down when we double the size of the system,” said Gupta.