The first annual OpenPOWER Foundation summit showcases new hardware and cloud services that utilize IBM's freely licensable processor technology, which adds developer-friendly changes.
IBM unveiled in August 2013 the POWER8 family of processors, and announced that the POWER architecture would be freely licensable by outside companies to produce processors and hardware based on IBM's design. The OpenPOWER Foundation, which was established to promote collaborative development of POWER systems, now has over 100 members (including market leaders such as Google, Canonical, NVIDIA, and TYAN) that are introducing new server and cloud solutions utilizing POWER8 processors.
The OpenPOWER Summit 2015 is being held within the GPU Technology Conference at the San Jose Convention Center March 17 - 19. Over two dozen presentations about the use of OpenPOWER are being given by academics and industry members, including people from Canonical, IBM, Rackspace, Oak Ridge National Laboratory, and Rice University.
Why is OpenPOWER important?
Presently, a vast majority of servers use Intel or AMD x86 processors; other competing architectures such as SPARC or MIPS are too specialized for general-purpose use, or too cost prohibitive to use. OpenPOWER takes the newest generation of the POWER ISA -- a mature platform for which previous iterations or derivatives have been used in a number of enterprise and consumer devices -- and opens that design to any foundry that wishes to make POWER processors, or other manufacturer that wants to produce hardware that utilizes POWER8.
The result of this collaboration provides a viable second source for performance computing -- though, that description alone is too simplistic. Because any foundry can produce POWER8 processors, OpenPOWER as a platform can more effectively compete on cost than architectures that rely on a single source. For a price comparison, according to a (paywalled) report by the Linley Group, the top-of-the-line 12-core POWER8 processor is estimated to cost $2,500, whereas the top-of-the-line Intel Xeon E5-2699 v3 can be found for $3,995.
However, the benefit of POWER8 isn't limited just to a lower cost-per-unit -- performance is a much more significant factor, with IBM claiming "up to 60% better performance per dollar spent." The design of POWER8 has higher memory bandwidth and multithreading capabilities in comparison to Intel's designs: Because of design limitations, Intel has not yet shipped a processor with more than two threads per core, whereas POWER8 processors have eight threads per core. Intel's 18-core Xeon E5-2699 v3 is limited to 36 threads, whereas the 12-core POWER8 has 96 threads.
Key design changes in POWER8 over previous iterations of the POWER ISA also boost performance in comparison to Intel Xeon. Memory buffering has been moved off the processor die onto external chips, each containing 16 MB of L4 cache and which connect to four DDR-3 1600 channels at 51.2 GB/s each, supporting up to 1 TB of DRAM at 205 GB/s peak bandwidth. According to the Intel Datasheet, the aforementioned Xeon processor has only four DDR4-2133 interfaces, limiting it to a peak bandwidth of 68 GB/s on all channels combined. Also of importance is the introduction of a complete little-endian mode on POWER8, which allows for easier porting of software originally targeted to Intel architectures.
What hardware or services use POWER8?
Taiwanese server manufacturer TYAN announced the first commercially available POWER8 server not produced by IBM, the TN71-BP012. It is a 2U server intended as an adaptable solution for any particular workload, and is a follow-up to its experimental, limited-release "customer reference system" from 2014. Interestingly, SoftLayer (which was acquired by IBM in 2013) will be deploying the Tyan TN71-BP012 as a part of its bare-metal cloud offering.
Rackspace is also prototyping motherboards for future bare-metal cloud offerings as part of its OnMetal service that utilize OpenPOWER, OpenStack, and Open Compute, for which Rackspace Senior Director Aaron Sullivan notes in a blog post that the benefit of deploying OpenPOWER as a bare-metal cloud offering is OpenPOWER "brings an increasingly open firmware stack, and deeper access to chips, memory, and storage than anywhere else."
IBM and Wistron are collaborating on a high-performance server codenamed Firestone, which they claim will reach speeds "five to 10 times faster than today's leading supercomputers," also utilizing technology from NVIDIA and Mellanox.
Verifiable security at the silicon level
In the Chinese market, a public-private partnership called the China POWER Technology Alliance (CPTA) was founded to adopt OpenPOWER as the platform of choice due to the openness and auditability of the design, as recent disclosures regarding security, and openly acknowledged features such as Intel Active Management Technology (AMT) have led to an abundance of caution in enterprise security.
CPTA member Suzhou PowerCore is producing its own POWER-based processors for the Chinese market, which will be shipped in a new line of two-socket servers called RedPower by Zoom Netcom.
What's your view?
Do you feel that the hardware features of OpenPOWER make it a viable potential migration path for your organization? Or, is your primary focus on ownership cost? Would migrating to OpenPOWER reduce the physical footprint of your on-premises data center operations? Share your thoughts in the comments.
Disclosure: James Sanders is an Associate member of the OpenPOWER Foundation.
- IBM's SoftLayer to use OpenPOWER servers (ZDNet)
- IBM wants enterprises to consider Power 8 (ZDNet)
- Is it time to consider Power 8 for Web workloads? (ZDNet)
- IBM spends $3 billion to push the far future of computer chips (CNET)
Note: TechRepublic, ZDNet, and CNET are CBS Interactive properties.