While AI (artificial intelligence) has been around since the 50’s, IBM was the pioneer in the latest AI cycle with their own custom solution dubbed Watson. Ever since the introduction of Watson and its ability to beat Jeopardy Champion Ken Jennings, the company has been increasing its investment in the space. IBM Watson is now an entire division of the company which indicates the importance they put on the future of AI.
Watson is only one part of IBM’s AI investment which I consider the “easy button” for those enterprise who don’t want to create everything from scratch. IBM also has DIY (do it yourself) infrastructure for cloud providers through POWER8, OpenPOWER, OpenCAPI, designed for cloud giant rolls their own AI software. But what about enterprises who are in the middle, those who want solid infrastructure and want to invest in the latest deep neural network frameworks? IBM announced an answer this week at SC16 and it’s called PowerAI.
IBM POWER8 and OpenPOWER for DIY public cloud AI
IBM intentionally designed the POWER8 architecture for future workloads like artificial intelligence, machine learning, and deep neural networks. In 2013, to expand the collaboration of hardware companies in the POWER domain, IBM established the OpenPOWER Foundation which allowed companies to closely collaborate on hardware and software to accelerate future workloads like AI and ML (machine learning). Part of the way that IBM has done that is with CAPI (Coherent Accelerator Processor Interface) and OpenCAPI interconnect for GPU and other forms of accelerators which dramatically improve specific workload performance. IBM’s POWER isn’t just for AI and ML, as POWER does well in HPC and accelerated databases like Kinetica, but those AI workloads run really well on POWER architecture.
IBM Watson, the enterprise AI “easy button”
As I said, for enterprises, IBM’s Watson has gotten some traction in certain verticals like healthcare and the “turnkey” solution makes it easier to deploy new ‘smart’ capabilities. However, there are still enterprises that want to develop their own software and might not want to use exactly the software that IBM is offering with Watson. They may want to develop with a specific deep learning frameworks like CAFFE, TORCH, TensorFlow or Theano. These are big credit card companies and drug research companies with the internal research arms and the budgets to roll their own sophisticated neural-network-based software.
PowerAI for enterprise DIY AI
For those companies, IBM has a new offering called PowerAI. PowerAI is a software toolkit with deep learning frameworks and building block software designed to run on IBM’s highest performing server in its OpenPOWR LC line, the IBM Power S822LC for High Performance Computing, which features NVIDIA NVLink technology optimized for Power and NVIDIA’s latest GPU technology operating at 80 GB per second.
The new PowerAI software toolkit and servers are a tightly integrated solution that allows enterprises to get into AI and ML quickly with a high-performance solution. IBM’s PowerAI hardware and software accelerate AlexNet and Caffe workloads at what the company says are “over twice the speed of the same solution utilizing IBM’s power versus an x86 solution.” The 2.2X figure came from IBM’s “Minsky” custom GPU Power8 accelerator server (4x NVIDIA P100) versus Intel Broadwell (4x NVIDIA M40) testing training time running AlexNet with Caffe, BVL Caffe and IBM Caffe at 50% accuracy. I did not test this myself nor did I watch the testing.
PowerAI has been in the planning phase for a while. I spoke with Sumit Gupta, IBM’s VP and head of POWER HPC and AI. He said the PowerAI solution is a culmination of years of work between the two companies who have been collaborating to tightly integrate their technologies together. I have seen that every year since the formation of the OpenPOWER Foundation it has yielded better and better results for IBM and their partners and PowerAI is the latest example.
IBM pioneered the most recent cognitive computing resurgence over the last decade with Watson, targeted at enterprises looking for the AI “easy button”. The company then followed Watson up with POWER8, OpenPOWER, CAPI and OpenCAPI infrastructure for AI public and private cloud implementations. IBM is now completing the AI circle with PowerAI for enterprises who want to use AI but have specific deep learning framework in mind. IBM gets enterprise, knows how to hold their hands, and I like where the company is going with PowerAI as it can only expand the footprint of AI, ML and deep learning. Nice job, IBM.