Share this post:
I have to tell you, after three days of intense deep learning sessions, I’m ready for a deep sleep. Luckily, NVIDIA CEO Jenson Huang knew how to start the day off right when he announced their next GPU, Volta. To put it lightly, Volta takes GPU-accelerated computing into the stratosphere, and we’re excited about how we can continue to leverage the unique CPU-to-GPU NVLink connection available only on the IBM POWER family of processors with this new accelerator.
NVIDIA Volta with NVLink 2.0 On POWER
As it is detailed in NVIDIA’s official Volta blog, Volta will feature NVIDIA NVLink 2.0 and IBM will continue to support the high-bandwidth interface between our Power family of CPUs and the Volta family of GPUs, and “NVLink now supports CPU mastering and cache coherence capabilities with IBM POWER9 CPU-based servers.” In addition, the Unified Memory technology in Volta GV100 includes new access counters to allow more accurate migration of memory pages to the processor that accesses the pages most frequently, improving efficiency for accessing memory ranges shared between processors. On IBM Power platforms, new Address Translation Services (ATS) support allows the GPU to access the CPU’s page tables directly.
And what I’m sure will be music to developers’ ears, PowerAI will support the NVIDIA Volta architecture announced today. For developers looking to leverage NVLink 2.0 on POWER9, this means the data transfer between the POWER9 CPUs and Volta GPUs is ten times faster than between Volta GPUs and x86 CPUs, which rely on the old PCI-e 3.0 interface, first introduced four years ago, and it is memory coherent, which makes programming GPU accelerators much easier for software developers by automatically moving data between the system memory connected to the POWER9 CPU and the GPU memory.
Visionary deep learning
Yonghua Lin, leader of the deep learning team at IBM Research China, also showcased her work around AI Vision, one of the key new features in the latest PowerAI update. AI Vision helps deep learning developers by simplifying the image recognition process in a simple GUI, allowing scientists to visually highlight and identify what they want the image recognition neural net to look out for.
As she explained, “Most organizations are facing challenges of no experience on DNN design, development, computer vision, or how to build a platform to support enterprise scale deep learning. This was the major motivation for AI Vision.” The hope is that this new feature will help organizations without a staff full of PhDs and data scientists begin to implement deep learning image recognition.
Out-of-core memory? Not a problem with POWER and NVLink
There’s a wide range of applications where the data model is too big for the device memory of the GPU, including deep learning training models as well as more traditional problems like HPC and simulation. IBM Research’s Rajesh Bordawekar from IBM’s TJ Watson Research Center sought to solve this problem in his Accelerated Analytics featured session.
This creates a lot of problems for researchers in deep learning and cognitive environments, as they’re often working with massive data sets that require immense computing power.
But thanks to NVIDIA NVLink’s CPU-to-GPU connection on POWER8, this becomes a thing of the past, as the unique memory bandwidth allows data to flow freely between the GPUs and the data store. The results of his work speak for themselves!
That’s a wrap!
As always, GTC has been an extremely insightful and fun conference, which highlights the best and brightest in deep learning. We hope that you’ve enjoyed our dispatches, and hopefully you’re pumped and excited about getting started with deep learning yourself!
You can download PowerAI for your Power Systems S822LC for HPC server and try PowerAI today on Nimbix’s deep learning cloud with NVIDIA NVLink and NVIDIA Tesla P100 GPUs.
If you are a data scientist using deep learning, we are looking for feedback on other software tools that we can add to make your experience better. Please post your comments below.