This article was written by Trond Bjerkvold, Senior Systems Consultant at Basefarm. It was originally published for Common Norway, a niched-interest organization for users of, among other things, the IBM Power servers.
AI is often discussed in widely different topics due to its multifaceted areas of usage. It will have an impact in education, health, finance, media, tourism and retail. AI entails everything from face recognition to self-propelled cars as well as GDPR analysis.
To solve artificial intelligence tasks, there are a number of open source software solutions. You can download and synchronize the software as needed, but this also entails a lot of time being spent on maintenance and updates.
Create a model – optimize efficiency
Once the software is in place, you can create a model based on a hypothesis and run data through it to achieve a result. The more complicated the model is and the more data you have, the longer it will take. The analysis is often based on pattern recognition, either from images, documents or other sources. Based on the results, you can change the model and run the data again. Requirements to enable this high performance are crucial.
The PowerAI combination means that you have software learning and deep learning software combined in a server/software package. Frameworks such as Caffe-bvlc, Caffe-ibm, Caffe-nv, Chainer, DIGITS, Torch, Theano and TensorFlow are already compiled and run on smooth updates.
Connect CPU and GPU – for a higher performance
IBM Power servers offer high performance, but they are not the common models we know from the AIX and IBM environments that are used by PowerAI software. PowerAI uses S822LC servers with NVIDA GPU and these only run Linux.
NVIDA is known for providing graphics cards for gaming PCs, while GPU features thousands of smaller processors on an adapter that makes Deep Learning tasks run faster. The GPU can be used as an application accelerator and can run a portion of the program code faster than the main processor. IBM and NVIDA have created a hyperlink, the NVLink, between the IBM Power Processor and NVIDA’s GPUs, so that the overall performance is much better than that of an x86 server.
FEWER SERVERS – FASTER RESULTS
In practice, this translates into having to maintain fewer servers, but getting the job done quicker. The new supercomputers used at Oak Ridge National Laboratory and Lawrence Livermore National Laboratory have IBM Power9 processor nodes with NVIDA GPU nodes.
For Oak Ridge, there are approximately 4600 nodes, each with 2xIBM Power9 and 6 NVIDA Volta GPUs. The previous supercomputer had over 18,999 AMD/NVIDA nodes. Even with far fewer nodes, the combination of Power9 and NVIDA really boosts performance: from 27-pet flops to 200-pet flops.
To run PowerAI, you have to download Ubuntu 16.04 for the IBM Power8, as well as the PowerAI package. These links can be found here.
The price for an S822LC (8355 GTB) with 16 Power8 cores, 256GB internal memory and 2 NVIDA GP100 GPU is around 350,000 NOK.