IBM Corp. is stepping up its deep learning game with an update to its popular PowerAI software for its IBM Power hardware systems.
The announcement came at Nvidia Corp.’s GPU Technology Conference in San Jose today. IBM developed PowerAI in concert with Nvidia, which makes graphics processing unit chips, releasing the inaugural version of the software last November at the SC16 supercomputing conference.
The software is essentially a toolkit for developers looking to leverage deep learning capabilities in so-called cognitive applications, which are used in self-driving cars, fraud detection systems and credit risk analysis, among others. PowerAI is laden with customized distributions of several popular machine learning frameworks, including Torch, Theano and Caffe, and is optimized to run on IBM’s Power Systems S822LC for HPC hardware, which is also designed for data-intensive, deep learning workloads.
IBM said its PowerAI systems are geared toward data scientists and developers working with these new machine learning frameworks, because they tend to be much more resource-hungry than traditional applications. As such, they have a tendency to quickly overwhelm systems based on the older Intel Corp. x86 chips.
The company’s Power Systems S822LC for HPC hardware provides extra oomph for these kinds of workloads by combining two Power8 processors with four accelerators from Nvidia’s Tesla P100 series. These chips are connected by Nvidia’s NVLink interconnector, which enables data transfer rates up to 12 times that of PCIe interfaces, which are normally used.
The updated PowerAI software platform comes with four major new features that were created to address some of the common challenges developers are facing, IBM said. These include a new AI Vision tool that’s designed to assist developers with little to no experience in using machine learning frameworks. The tool allows developers to begin training and deploying deep learning models almost instantly, IBM said. In addition, Power AI now integrates with IBM’s Spectrum Conductor cluster virtualization software, so that structured and unstructured data sets can be prepared for deep learning training more easily, the company said.
There’s also a new tool called DL Insight that the company says can be used to get better accuracy from deep learning models. The tool works by moniroting the deep learning process in real-time, and adjusting paramters to optimize the entire system’s performance.
Perhaps most significant, PowerAI comes with a newly updated, distributed version of TensorFlow, the popular open-source machine learning framework that was developed by Google Inc. IBM said its customized TensorFlow distribution can leverage virtualized clusters of GPU-accelerated servers to massively accelerate deep learning training processes. As a result, it’s possible to reduce training time from weeks to just hours, IBM reckons.
“Power AI reduces the frustration of waiting and increases productivity, Bob Picciano, senior vice president for IBM Cognitive Systems, said in a statement. “Power Systems were designed for data and this next era of computing, in contrast to x86 servers which were designed for the client/server programmable era of the past.”