06 Sep IBM’s OpenPOWER: A Lot Has Changed In Two Years
The OpenPOWER consortium was born of IBM’s own volition only two years ago last month. Back then, OpenPOWER was a partnership between five companies, IBM, Google, Mellanox Technologies, NVIDIA and Tyan. Recently, the OpenPOWER Foundation marked their two year anniversary by talking about the growth of the foundation and its membership to 149 members in 22 countries and the broad licensing that has occurred thanks to IBM’s open licensing model. I talked about the OpenPOWER Foundation back in April of last year and the announcement of IBM’s new Power8 processors with their OpenPOWER partners. Since launch and last April’s announcement, OpenPOWER has really come along quite a bit and I wanted to take a 50,000 foot level view at where OpenPOWER started and where IBM and the OpenPOWER Foundation have gone since then.
There were two sides to OpenPOWER’s launch two years ago, one on IBM and POWER8 systems and another for the OpenPOWER Foundation’s launch where they wanted to create an open, server ecosystem. The launch two years ago answered many questions on IBM’s POWER8 systems paired with IBM software for IBM’s enterprise customers, of which I was complementary. I was also quite impressed with their accelerator interface, called CAPI (Coherent Accelerator Processor Interface). If you weren’t a believer in CAPI or accelerators before, Intel’s $16B Altera acquisition should close the debate. However, there were many unanswered questions about systems outside of the IBM enterprise and software sphere, particularly in the hyperscale server market.
The hyperscale server market over the years has become a DIY (do-it-yourself) market with Google, Amazon.com AWS, Microsoft Azure, Baidu and FaceBook leading the charge. When OpenPOWER launched, I thought hyperscale was a not a place where IBM had recent history or a business model to attack in the future.
Since then, OpenPOWER has gained a lot of momentum. Google, the hardware they showed off a few weeks after the OpenPOWER launch, and the commitment to port their software for at least a POC (proof of concept) was the first big hyperscale milestone. OpenPOWER then had a strong showing at Open Compute versus ARM Holdings. In two years, IBM and the OpenPOWER consortium have achieved a lot including support for all Linux distros. 2015 as a whole has been a very busy year for OpenPOWER’s work groups with all of them completing their charters and many of them releasing their specs. Some of these include the CAPI Linux SDK, 64b ABI (Application Binary Interface), 25g IO Spec, OpenCL SDK and many others. The Foundation is up to 11 chartered workgroups and they say they tell me they have more on the way.
In addition to the working group announcements, there have been big announcements from OpenPOWER members in different ways. Even though Altera will now be part of Intel (pending government approval), IBM and Altera did work together to create a DMI connection between Altera’s FPGAs and IBM’s Power8 CPU. Convey created a developer kit which is based on the company’s Xilinx-based co-processors and IBM announced with Suzhou PowerCore, the first Chinese POWER licensed chip, the CP1. Suzhou PowerCore announced this summer that they reached general availability for the CP1 processor. I’m looking forward to third party benchmarks on the processor and accompanying systems. There were also announcements from companies like Tyan who released the first commercially available OpenPOWER third-party server and a CAPI-based interconnect solution from Mellanox as well. There were also new systems and platforms from the likes of Rackspace Hosting, RedPower, Inspur, Wistron and Cirrascale.
Many of these developments show IBM and OpenPOWER’s focus on China. China’s government-led operations in finance, military, research, healthcare and education have a government mandate to have all new systems on 100% controlled hardware, software and middleware. This is referred sometimes as “China 2016” in the industry and OpenPOWER wants the datacenter side of it to be based on their technology.
OpenPOWER has also shown its relevance in the HPC (high performance computing) space in ways that IBM has struggled in the past. Just this year, IBM, Mellanox and NVIDIA won two DoE (Department of Energy) supercomputer bids for Oak Ridge and Lawrence Livermore on tap in 2017. All three companies are founding members of the OpenPOWER Foundation and collaborated closely to win the $325 million deal. These systems are expected to reach into the 150+ Petaflop space, which is at least 3x faster than anything today, including China’s Tianhe 2. This is potentially a huge boon for IBM’s HPC business because it not only lends credibility, but it also lends expertise. As the world moves towards larger and larger computing problems a need to have HPC expertise will be important for any high performance solutions provider or chip vendor. It also helps that IBM, NVIDIA and Mellanox have synced up their roadmaps for HPC in order to bring vast improvements with each year.
IBM isn’t doing all of this just to be nice- the point of OpenPOWER and creating new CPU architectures is to make money. IBM looks to make money in a multitude of ways, but there are three primary methods, selling cloud and scale out processors as well as HPC solutions that use their processors. As for the relevance of HPC versus the other datacenter markets, Jimmy Pike, Moor Insights & Strategy technologist residence commented that, “And if you don’t think HPC is that commercially relevant, you haven’t looked under the hood of a big data solution….”
IBM also plan to make money through collaboration and enablement services, which IBM always looks for ways to implement. Last but not least is the IP part of things which includes IP licensing and royalties from the various licenses and agreements that are created as a result of OpenPOWER’s growth. Finally, let’s also not forget that any IP development into OpenPOWER is money IBM doesn’t have to expend, and quite frankly, couldn’t shoulder on their own.
IBM has come quite a long way over the course of the last two years, and it shows. They have some pretty significant HPC deals and there’s no saying what kind of demand Watson could create for IBM’s POWER Systems and the OpenPOWER Foundation. However, the OpenPOWER Foundation still has a lot to prove by ways of non-IBM silicon, non-IBM systems and non-IBM cloud software benchmarks since much of what’s been shown has centered on IBM, but there are others that are part of the Foundation that should have something more to contribute. 2014 and 2015 were building years for the OpenPOWER Foundation and 2016 will be the proving ground as systems move from research to POC (proof of concept) to mass deployments. Keep in mind, too, in this hyperscale space, Intel has nearly 100% of all the processor business and are architecting business models and architectures to combat this. In HPC, Intel just closely partnered with Hewlett-Packard on a multi-year HPC collaboration effort.