By Joe Clabby, President, Clabby Analytics
In August, 2013, IBM announced the formation of the “OpenPOWER Consortium” which formally became the “OpenPOWER Foundation” in April, 2014. The goal of this foundation from the outset has been to create a vibrant, open ecosystem around IBM’s POWER microprocessor architecture – enabling customers and systems/software vendors to build their own systems designs using IBM’s POWER8 and successive generation POWER-based architectures. To help expand this ecosystem, IBM announced that it would make processor specifications, firmware and related software available to the evolving OpenPOWER community – essentially making the POWER processor an open standard.
As could be expected with any standards organization, initial progress was slow. The foundation had to create a structure and organize an operating committee, formulate strategies, set goals, create technical committees, build reference designs, create by-laws and operating procedures, document various interfaces – all of the usual procedural collaborative activities that needed for bringing new POWER-based solutions to market.
The OpenPOWER Summit 2015
Despite those challenges, the OpenPOWER Foundation was able to help members focus on bringing POWER-based solutions to market. In March, 2015, (less than a year after the OpenPOWER Foundation opened its doors), the OpenPOWER Foundation was ready to share its activities, goals and new product offerings with the marketplace at the first OpenPOWER Summit (the OpenPOWER Summit 2015 held in San Jose, California). At that time, the OpenPOWER Foundation had recruited a critical mass of customer and vendor members (approaching almost 80 members at that time) – and these members had either individually or collaboratively created several new POWER-based systems components and designs. The most notable new server designs included TYAN’s TN71-BP012 POWER8 server; the Cirrascale 1S Power8/NVIDIA GPU Server; the ChuangHE 1U server; and, IBM’s own S824L POWER8/NVIDIA GPU Server. Rackspace and Google also showed new planar prototypes. Additionally, other OpenPOWER Foundation members showed new coherence attached processor interface (CAPI) and FPGA adapters. I remember being particularly impressed with the new accelerated systems designs that featured FPGAs and GPUs working with POWER processors to accelerate analytics performance. I detailed several of these new designs shortly after having attended the OpenPOWER Foundation Summit 2015 in this blog.
The OpenPOWER Summit 2016
On April 6th, 2016, I attended the OpenPOWER Foundation Summit 2016, again held in San Jose, California – and again the OpenPOWER Foundation was able to demonstrate great progress in building an open, collaborative ecosystem around the Power8 architecture. The list of 59 new products that had become part of the ecosystem included FPGA cards, chip liquid coolers, memory cards, high-speed switches, new POWER-based system designs, new boards and expansion kits, GPU accelerators, and even a new POWER chip for the China market. Further, OpenPOWER Foundation membership had grown to more than 200 members.
The biggest eye-opener at this Summit was a project being jointly developed by Google and Rackspace that will create a new open server specification based on the forthcoming POWER9 architecture. Both companies plan to submit a candidate server design to the Open Compute Project (a Facebook-inspired initiative that allows vendors and users to share advanced data center and server designs with the information technology [IT] industry).
The Google representative, Maire Mahony, a hardware engineering manager, indicated that the company’s server deployment decisions are highly influenced by total cost of ownership (TCO) and by performance – which is why Google is willing to entertain using POWER-based servers with the Google infrastructure. She also indicated that Google expects that their new POWER9-based design that will feature the use of 48 volt architecture reduces power distribution losses by up to 30% compared to 12 volt systems, while delivering significant improvements performance for certain applications. As for Rackspace, the company showed its “Barreleye” Power-based cloud server which has moved from the lab to the data center – and should be broadly available later this year. The Barreleye server was formally accepted by the Open Compute Project in February 2016. Google and Rackspace both see benefits in using POWER9-based servers for particular workloads, and plan to continue to work together on future Open Compute Project designs.
The reason that I consider the Google/Rackspace collaboration to be a big “eye-opener” is that I have long argued that POWER architecture should be more widely deployed in large clouds. There are huge processing efficiencies to be gained by using POWER processors instead of Intel processors for many types of workloads. POWER processors can process twice as many threads as their Intel competitors, they offer more cache, and new interfaces such as CAPI can deliver lower cost storage/memory connections with far less latency. I have long been surprised that POWER processors don’t dominate more industry clouds – and attribute the lack of adoption of POWER-dominated clouds to tardy development and promotion of Linux on POWER. If companies like Rackspace and Google are starting to open the door for POWER-based servers – and if they can prove processing and energy efficiencies – it may not take long for other large cloud vendors to get the message.
The POWER9 Discussion
Other big news at this year’s summit included a preview of the next generation POWER9 architecture (in very general terms) – as well as a roadmap of priorities for the OpenPOWER Foundation.
The POWER9 discussion lasted only a few brief minutes, but indicated that POWER9 would offer:
- 24 newly designed POWER9 cores that would levereage execution slices for improved performance on cognitive, analytic and Big Data applications;
- Large, low-latency eDRAM cache for big datasets;
- Cloud-focused innovations for energy efficiencies, security and Quality-of-Service;
- State-of-the-art I/O subsystem using PCIe Gen 4; and,
- New and improved interfaces for hardware acceleration including high-bandwidth, GPU interconnect using the evolving NV link2.0 standard; next generation CAPI 2.0; on-chip compression and cryptography accelerators and a new 25Gb/s advanced accelerator attach bus.
IBM also announced that the 1st chip in the POWER9 family would be optimized for 2 socket scale out servers and hyperscale datacenters, as well as offering DDR4 direct attach memory channels. As POWER9 further matures, it will be designed to address a broader range of scale out and enterprise servers.
I expect that IBM will provide the analyst community with a far deeper briefing as the release date for POWER9 approaches – but for now what these specs tell me is that the designed-for-data POWER architecture is about to be turbocharged.
The OpenPOWER Foundation Roadmap
As could be expected, the OpenPOWER foundation will continue to focus on member growth in order to expand the ISV and developer ecosystem for POWER-based servers. The foundation also wants to create more solution stacks; increase its market visibility; and focus on certain other initiatives such as deep learning at scale. The foundation also plans to take its show on the road, offering summits in Europe and China. As it does this, expect a big part of the messaging to focus on what the foundation believes that it is delivering: a forum for collaborative innovation.
At this year’s summit, I observed that the OpenPOWER Foundation has made great progress in working its way out of the “set-up doldrums”. Assembling a standards organization, related procedures and processes, and organizing personnel to help drive momentum on a large scale has had to have been a daunting task. But with a lot of that rigmarole out of the way, the Foundation is now better able to focus on what is truly important – the creation of solutions that drive value for customers.
The way I see the POWER architecture at IBM is two-fold:
- IBM, a master designer of microprocessor architecture and creator of advanced system designs will continue to build POWER Systems for its focus markets (cloud, analytics, mobile and security) – frequently using IBM hardware and software taken from its own portfolio in order to assemble new POWER-based solutions;
- The OpenPOWER ecosystem will address the needs of buyers who need more custom-type solutions – or that need high-powered designs that are not part of IBM’s focal strategies. The ecosystem enables vendors to collaborate and innovate – and bring other ideas and implementations to market.
In each case, IT buyers and users win. The traditional Power Systems group has now reported full year growth of +4% in constant currency. And in Linux, the Power Systems group saw a 4X growth rate year-to-year. IBM’s Power Systems organization is growing and moving profitably in the right direction.
The right products and strategies are helping IBM customers capitalize on more powerful, POWER-based Linux servers. Meanwhile, the OpenPOWER ecosystem has a dozen very exciting projects including several new accelerators (as described in last year’s blog) as well as new communications off-loading facilities being developed at the network level. These projects will result in faster, more efficient systems – again delivering major performance and efficiency advantages to the end-user community.
In summary, the OpenPOWER Foundation has made great progress in delivering innovative new system designs and components since last year, and it has significantly grown its member base. Next year I want to see more members – and more concrete customer solutions with real world proof points on efficiencies gained using POWER-based hardware and software solutions.