With IBM’s pending sale of it’s x86 server business to Lenovo, their remaining server business is squarely focused on Power, and that holds great promise for innovation via the OpenPower Foundation. The concept of an “open” CPU architecture is not new. In 2006, Sun Microsystems released the complete design of its UltraSPARC T1 processor to OpenSPARC.org. Oracle’s acquisition of Sun in 2010 ensured that effort didn’t continue, although OpenSPARC is still used in a few university courses. With 26 members and counting, including the likes of Google, NVIDIA, Mellanox, and Samsung, the OpenPower Foundation appears set for greater success, as demonstrated in today’s OpenInnovation Summit.
As part of its work with the OpenPower Foundation, NVIDIA is adding CUDA software support for NVIDIA GPUs with IBM POWER CPUs. IBM and NVIDIA are demonstrating the first GPU accelerator framework for Java, showing an order of magnitude performance improvement on Hadoop Analytics applications compared to a CPU-only implementation. NVIDIA will offer its NVLinkTM high-speed GPU interconnect as a licensed technology to OpenPOWER Foundation members.
GPUs with NVLink are the perfect accompaniment to OpenPower. There was plenty of discussion at today’s OpenInnovation Summit about the powerful memory subsystem of Power8. Even less powerful memory subsystems available today overtake the fastest PCIe interfaces to the GPU. NVIDIA has worked closely with IBM on roadmaps and NVLink promises to keep up with Power8 and beyond memory systems, ensuring GPUs on OpenPower systems will be able to take full advantage of all the CPU memory bandwidth. In turn, this will enable OpenPower systems to realize the full potential of the Unified Memory architecture introduced in CUDA 6.
IBM has a built-in market for future Power + GPU systems with its own software divisions. At the GPU Technology Conference earlier this year, IBM demonstrated some of its commercial software applications written in Java running on a Power + GPU test server. Java applications typically have many parallel threads that can be accelerated through the use of GPUs. And of course CUDA based code, which executes on the GPU, is basically unchanged no matter if that GPU is attached to a Power, ARM, or x86 based CPU.
With the OpenPower foundation, IBM appears to have taken the motto of innovation happens everywhere to heart and already has a good start on driving adoption. As I mentioned in my last blog, there are already companies in China working on OpenPower designs. It is great to see innovation in the processor and server world is alive and well, promising to give IT customers more choices and greater value for well into the future.