s published in insideHPC
IBM’s announcement this month that it would be opening up its Power processor IP for licensing by third-party developers offers hope that the future of
the data center will not be just a two-way shootout between x86 and ARM. To that end, IBM launched the OpenPOWER Consortium, a group devoted to bringing
together the hardware, system software, and application resources for would-be Power developers. Besides IBM, the initial consortium members are Google,
Nvidia, Mellanox Technologies, and the Tyan Computer Corp.
The IBM initiative mirrors the business model successfully pursued by ARM Ltd., which that company has used to dominate the consumer electronics space
for hand-held devices like tablets and phones. With its new 64-bit processor design, ARM now has its eye on the PC, workstation and server market.
Although ARM Ltd. is in the sole business of licensing its IP, IBM is in it to expand the ecosystem for its Power-based portfolio of servers, software
IBM is responding to a rapidly changing competitive landscape that is both steadily marginalizing proprietary IT technologies, and subsuming component
technologies into larger more capable products (i.e. systems on a chip [SoC]). Thus the company is extending its power strategy to attempt to better
compete in this brave new world of computing. By further opening up the architecture IBM is positioning itself to gain significant new share in
the market, in which case the company will be in position to benefit from software, services, and yes, even hardware sales.
x86 has steadily eroded server market share of all RISC-based CPUs, like Power and Sparc, a trend that is mirrored in the HPC submarket. According
to our latest research (2013), RISC processors are deployed in less than 5 percent of all HPC systems, with x86-based systems commanding 94 percent
of the market.
Putting the Power architecture on a more even footing with x86-based server platforms and future ones based on ARM will be an uphill battle. Those two
legacy architectures are already well-established in their markets and are now poised to do battle on each other’s turf. Can Power challenge them in
the data center? It’s too early to tell, of course, but there a number of reasons to be optimistic:
- It’s IBM. The company has vast resources and an impressive set of system software and applications devoted to the Power ecosystem,
and much of that can now be leveraged by external partners. Although theoretically IBM could lose potential market share for Power servers, the
company is not likely to be directly challenged for their current product set, since it sells integrated systems. In any case, IBM would prefer
to sell more of its high-margin software and services for Power systems, rather than just the hardware.
- x86 Fatigue. The industry is wary of closed architectures, especially those dominated by a single vendor — in this case, Intel.
The more successful x86 becomes in any given market — and it has been remarkably successful in the server space — the more leery users
become. Intel has also been making moves into building systems (servers, motherboards, system management, network components, etc.), a development
unlikely to endear the chip maker to its OEM partners. An open architecture is a much more attractive model for users and suppliers.
- Power Differentiation. The architecture spans a range of usage models — everything from supercomputers to low-end servers and
workstations — and at the high end can outperform x86 CPUs on a per core basis. The upcoming Power8 platform, which will be the first design
available under the open license, will include such features as improved simultaneous multithreading (SMT), transactional memory, bigger caches,
and support for FPGAs and accelerators. It will also add support for PCIe 3.0 as well as a technology that rides atop it, known as Coherently Allocated
Processor Interface (CAPI), to provide coherent memory support between the CPU and coprocessors.
- Google. The addition of Google to the OpenPOWER Consortium implies the search giant sees a future for the architecture in its own
business, most likely in data analytics, but possibly in other areas as well. Since Google tends to deploy computing systems in acreage-sized allotments,
this could bode well for Power’s future. And if Google thinks Power is a worthwhile investment, so might its competitors.
- Networking and Storage. Although a lot of the focus has traditionally been on compute servers, networking and storage appliances are
also a large market for commodity processors and these tend to have less dependency on legacy system software. With the emphasis now shifting to
SoCs for compute, networking and storage, a licensable Power architecture could be very attractive to vendors building gear powered by custom processor
- GPUs on Power. Nvidia’s involvement in the consortium points to what could be an important collaboration between IBM and Nvidia. According
to the press announcement the two companies “will work together to integrate the CUDA GPU and Power ecosystems.” That combination would deliver
a unique set of capabilities to the market, and one which Nvidia sees as a way for the Tesla products to shed their x86 dependency as well as break
out of the HPC niche into the broader (and faster growing) big data market. For IBM, it offers a path to build x86-free servers with top-of-the-line
The Nvidia connection is all the more interesting considering the GPU maker’s recent acquisition of PGI, which brings deep compiler expertise, and particularly
CPU compiler expertise, in-house. With Nvidia’s commitment to ARM and now Power, the company is strategically aligning itself with technologies that
free itself and its customers from x86-based systems.
What is missing from this initial announcement is any indication that a processor, server, storage, or network vendor is ready to license the technology
and build their own Power processors. Although it’s plausible that well-known chip makers such as AMD, Fujitsu or Samsung could consider such a
license, a more plausible first mover might be a Chinese server or processor vendor with a particular aversion to proprietary technologies. Likewise,
other European or Asian vendors might see this as a way to diversify their server portfolios.
Overall, IBM’s open strategy appears sound; it responds to the industry dyamics around open technology and coopetition, while maintaining the company’s
ability to cross-sell software services. In addition, with the trend toward using customized SoC designs for servers, storage, and networking, it makes
sense to grow the market by expanding the ecosystem — that is, allowing multiple companies to support the processor. It remains to be seen, though,
if there is an appetite for yet another industry-standard processor for the data center.