OpenPOWER Foundation At Two - IBM-Lead Ecosystem Is Moving Fast

Beyond the Google and Rackspace “Zaius” server rack design based on the Power architecture, the second annual OpenPOWER Summit demonstrated progress across most aspects of the OpenPOWER Foundation’s budding new data center ecosystem.

There are two ways to interpret OpenPOWER momentum. Both are meaningful from different perspectives.

The first interpretation is simply that the cloud data center market is still looking for a viable alternative to Intel ’s Xeon processor line. Xeon products now command all but a scant few percent of all server unit shipments. As good as Xeon products are, markets abhor monopolies and Xeon now holds by anyone’s definition a monopoly in server processors. The server market accepted x86 as a de facto standard when AMD’s Opteron processors played a strong second to Intel’s Xeon, but AMD performance faltered and the company is now waiting for their Zen cores to effectively re-enter the server market in 2017. This has left a gap for the last few years that Intel has exploited to reach over 97% share of all server shipments in 2015.

The second interpretation of the OpenPOWER Foundation’s momentum is that new market dynamics are propelling interest in OpenPOWER as a high-performing competitor to Intel’s Xeon E5 and E7 class servers – the convergence high performance computing (HPC), big data analytics, and advanced business intelligence (BI) workloads, along with new deep learning and machine learning techniques.

Viable Alternative

IBM appears to have borrowed their OpenPOWER leadership philosophy of benign dictatorship from Linux. I say “benign” because both are authoritarian stewardships intended to grow their respective ecosystems. In the case of OpenPOWER, IBM will tightly control OpenPOWER partner implementations of IBM’s POWER8 and POWER9 instruction set logic.

The result is that OpenPOWER partner processors and system-on-chip (SoC) designs will share the same performance behaviors, therefore software optimized for any of the OpenPOWER partners’ chips will be optimized for all of the others. This is in contrast to ARM’s architecture licensees, each of whom implements a different microarchitecture with different software optimization behavior from the rest. Both Google and Rackspace mentioned the effort needed to optimize applications across processor and SoC products in an instructions set ecosystem as a key factor in their decision to invest in deploying OpenPOWER based servers instead of ARM based servers.

The OpenPOWER ecosystem’s biggest challenge is that non-IBM SoCs are for the most part unavailable. IBM signaled at the OpenPOWER Summit that partner chips won’t be on the market until 2018. This means that IBM’s own POWER8 and POWER9 processors will be the only OpenPOWER processor products on the market for perhaps the next two years.

The only publicly identified non-IBM OpenPOWER SoC under development is the Suzhou Powercore’s CP1 processor that powers Zoom’s RedPower C210 server, but neither CP1 nor C210 have been seen running outside of China.

OpenPOWER ecosystem appears to be rapidly maturing, despite lack of partner OpenPOWER processors

At the summit, the OpenPOWER Foundation announced their “OpenPOWER Ready” system certification program, which reinforces IBM’s instruction set control (specific processors and firmware) and un-core options (boot management and power/thermal devices).

The OpenPOWER Ready v1.0 specification states:

• At least one [IBM external] memory buffer with memory for POWER8, POWER8 with NVIDIA NVLink, or CP1
• Firmware to initialize the hardware is a modest derivative of the OpenPOWER Abstraction Layer (OPAL) FirmWare base, with no firmware API/ABI changes that have not been accepted upstream
• A system level XML file as input to the OpenPOWER Build process for host firmware
• A boot management chip to manage the Power On Reset (POR) and initial Initial Program Load (IPL) is needed
• [Specific types of] Thermal and Power devices

Even with self-reporting, this is a much more detailed specification than the Open Compute Platform (OCP) Inspired and Accepted certification levels.

Higher Performance

IBM is considered a leader in HPC. However, over the last few years the company’s influence in the supercomputer TOP500 list has waned – PowerPC-based IBM Blue Gene/Q accounts for just 3.8% of systems on the most recent November 2015 list, and POWER7 accounts for another only 1.4%. It is clear from this year’s OpenPOWER Summit that IBM is attacking the scale-out HPC market through the OpenPOWER Foundation with POWER8 performance and memory bandwidth. OpenPOWER also supports a variety of compute accelerators, notably NVIDIA’s newly announced flagship P100 (Pascal) GPU compute module. With IBM placing big bets on Its Watson technology and machine learning, this is not a surprise.

Inventec , Wistron and Zoom showed OpenPOWER systems using IBM’s POWER8 processor and incorporating multiple NVIDIA P100 modules. In the photos below, the large covered connectors are the NVLink connectors – where each P100 module plugs into two adjacent NVLink connectors. Wistron’s system (marked C) shows copper heat sinks on two populated P100 modules

I found it noteworthy that IBM and the OpenPOWER Foundation are investing in scale-out P100 designs. It is an indication that IBM sees more value in focusing on the processor’s instruction set and is willing to let its partners focus on application specific acceleration. Competitively:

• Intel is focusing on its own Xeon Phi accelerator roadmap in order to control more of the silicon share of wallet in HPC systems
• QCT has opportunistically productized the same Xeon-based 8x P100 motherboard used in NVIDIA’s own DGX-1 developer system to offer their customers a choice between Intel’s Xeon Phi and NVIDIA’s P100 accelerators.

In addition to GPU-based acceleration, the OpenPOWER Summit highlighted several FPGA accelerator solutions plus Wistron’s “Dark King” pre-production POWER8 based 4U/4S large memory design. Dark King implements 16 memory mezzanine cards with eight DIMM slots each for a total of 128 memory DIMMs. Using 8GB DIMMs Dark King can house 4TB of main memory, or 1TB per POWER8 processor socket. The Dark King design uses CAPI-attached network interfaces, which will be available at 10Gbit to 40Gbit and potentially faster. Wistron’s design target for their Dark King design is in-memory databases.

Conclusion

The OpenPOWER ecosystem appears to be rapidly maturing, despite the lack of non-IBM OpenPOWER processors. My conclusion is that either way you cut it – price foil against Intel Xeon processors or worthy HPC competitor to Intel’s E7 class Xeon processors plus Xeon Phi (Knights Landing) – IBM’s OpenPOWER strategy has merit. The real litmus test will be volume cloud and HPC deployments, and for that we’ll have to see how 2016 progresses. A newly announced Google defined POWER9 design can’t hurt POWER8 based OpenPOWER system sales, In fact, it may drive other public and private clouds to evaluate POWER8 sooner rather than waiting to catch up to Google when POWER9 systems ship.

-- The author and members of the TIRIAS Research staff do not hold equity positions in any of the companies mentioned. TIRIAS Research tracks and consults for companies throughout the electronics ecosystem from semiconductors to systems and sensors to the cloud. IBM sponsored the author’s travel to the OpenPOWER Summit.