4.3 KiB
| title | date | categories | tags | |||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Exploring the Fundamentals of OpenPOWER, POWER9 and PowerAI at the University of Reims | 2019-06-25 |
|
|
By Professor Michaël Krajecki, Université de Reims Champagne-Ardenne
Last month, the University of Reims hosted a workshop introducing the fundamentals of the OpenPOWER Foundation, POWER9 and PowerAI. Students and faculty from the University were joined by experts from IBM POWER Systems, EBV Elektronik and the Barcelona Supercomputing Center for a great session!
Multiple topics relating to POWER9, deep learning and PowerAI were discussed.
-
- Thibaud Besson, IBM Power Systems: Fundamentals of OpenPOWER Foundation, POWER9 and PowerAI: Besson discussed why POWER9 is the most state-of-the-art computing architecture developed with AI workloads in mind. He also showcased PowerAI, the software side of the solution, explaining its ease of use and straightforward installation that reduces time to market for implementors.
-
- Franck Maul, EBV Elektronik: On Xilinx Offerings: Maul presented Xilinx products that are going to revolutionize the AI market in the near future, explaining why Xilinx’s offering is the best fit for customers in the current market. He also showed off Xilinx FPGAs, emphasizing their perfect fit with IBM AC922 servers.
-
- Dr. Guillaume Houzeauk, The Barcelona Supercomputing Center: How Fluid Dynamics Can Be Implemented on POWER9 and AC922 Servers: In one of the day’s more technical sessions, BSC examined how a major Spanish car manufacturer has implemented Fluid Dynamics in a cluster of AC922 servers to improve automotive design and to reduce product cost and cycle time.
-
- Ander Ochoa Gilo, IBM: Distributed Deep Learning and Large Model Support: Ochoa Gilo dove into the benefits of deep learning, showing not only how we can overcommit the memory of the GPUs both in Caffe and Tensorflow, but also how to implement it. Using live examples, Ochoa Gilo explained how deep learning is accelerated through AC922 servers, allowing users to select images with up to 10x more resolution vs x86 alternatives.
He also demonstrated another useful feature of PowerAI, distributed deep learning, which allows a model to be trained on two servers using RDMA connectivity between the memory of the AC922 servers, reducing training time. Finally, Ochoa Gilo showcased the SnapML framework, which allows non-deep learning models to be accelerated by the GPUs, reducing the training time by 4X. He ran live examples that demonstrated its effectiveness right out of the box – some researchers in the room were so impressed by the framework that they implemented it in their clusters before the demonstration ended!
-
- Thibaud Besson, IBM POWER Systems: PowerAI Vision, CAPI and OpenCAPI Interface to FPGA on POWER: Thibaud Besson returned to explain why PowerAI Vision is a fundamental solution for companies that cannot afford to hire the world’s best data scientists. In a live example, he created a dataset from scratch, ran a training and then put it into production. The dataset was able to be monetized in minutes, offering the usefulness of the model to any software that can make an API REST call.
To wrap up, Besson explained the usefulness of being an open architecture, diving into CAPI and OpenCAPI and the benefits of using it in I/O intensive workloads.
AI is a key topic of interest for the University of Reims and its partners as further projects out of the University explore AI in agriculture and viticulture. As such, participants learned more about OpenPOWER and AI, and speakers in return were able to better understand the needs of our local researchers. All in all, the workshop was well-received and highly engaging. Thank you to everyone who participated!