With the continuing sales of consumption devices like tablets and smartphones, demand is increasing for the servers that feed them. The current consensus is that high-end enterprise type servers are a stable market, not a growth market, while the mid-range and low-end servers will continue to grow. While Intel (NASDAQ:INTC)-based servers have taken the lion's share of the server market in the past several years, competition is intensifying as Intel's last two processors failed to improve top-end performance considerably, while focusing more on performance at lower wattages. Many look to AMD (NASDAQ:AMD) as a potential competitor with its Zen and K12 releases in 2016, or pay attention to Qualcomm's (NASDAQ:QCOM) intentions in the server market with interest. While it is important to never discount companies, one has failed to have a competitive server processor for nearly a decade, the other has no experience developing this class of processors.
A company that already has a processor with significantly greater performance than Intel's, is IBM (NYSE:IBM), and its POWER8 line. Many like to think of Intel as a top-shelf provider of server hardware, whereas it's really more of a mid-range brand that falls short of IBM at the highest end of the segment. IBM has historically been very proprietary with its line of processors, while Intel has been more open, and less expensive, while still offering good enough performance for many scenarios. For IBM to gain greater adoption for the POWER line of processors, it has to leverage its superiority in security, performance and scalability, while eliminating the previous impediments to POWER adoption. The OpenPOWER initiative is this attempt, and offers far greater performance, while being more open than the x86 platform. But, where the price falls is an important consideration we have yet to see, although Tyan's reference platform appears to be reasonable at $2700.
My intention with this article to familiarize the reader with OpenPOWER, because it seems to be given far too little attention by the press, and too little consideration when looking at the markets going forward.
It all starts with the POWER8 family of processors. POWER8 does not attempt to be cute like Intel's Haswell and Broadwell; it's a brute, and doesn't want to be in tablets, or laptops. It doesn't want to be in low-power desktops. It's a server chip, and it wants to dispatch heavy workloads, and feed the consumption devices Intel pays companies to put its processors in. At one time, Intel did try to compete with POWER at the high-end, with the Itanium line, but its failure there is well documented, and has forced Intel to try to use the same chips it wants to throw in tablets, into high-powered servers. These compromised chips still do an adequate job for many workloads, but in others are less than ideal compared to higher-end chips and platforms with a more singular purpose.
I attempted to get an accurate comparison between Intel server processors and IBM's using the industry standard on SPEC CPU2006 benchmarks. POWER8 results were only available on SPECint*_rate2006, and SPECfp*_rate2006, so I used those two. IBM's E880 was only tested in an 8-socket configuration, so I used that as a standard, against 8-way Intel processors. I wanted to see the results for both per core performance, and per socket performance. For that, I chose IBM's E870 for per socket performance and E880 for per core performance. For Intel, I chose the E7-8890 v2 for the performance per socket, and E7-8893 v2 for per core performance. All represent the best processors in their respective categories, for each company, that had available results.
Intel's E7-8890 v2 has 15 cores, whereas IBM's E870 only has 10, so the per socket performance is where Intel shines. Well, kind of, they still finished in second place, out of two, with a score of 4570, or 571.3 per socket in SPECint*_rate_base2006. Although Intel's processor has 50% more cores, IBM's E870 finished with a score of 4830, or 604 per socket on the same benchmark. In floating point, Intel suffered at the hands of the mighty E870, scoring 3240 points, or 405 per socket, in SPECfp*_rate_base2006. By contrast, the E870 scored 4500, or 562.5 per socket.
For per core, as mentioned, I used the illustrious E7-8893 v2, a six core processor, for Intel. For IBM, I used the E880, an eight core processor. The E7-8893 v2 scored 2290 for SPECint*_rate_base2006, or 47.7 per core. IBM's E880 scored 4170, or 65.2 per core. In floating point, ugly becomes hideous for Intel, as the E7-8893 v2 scores a paltry 2000, or 41.7 per core. IBM's redoubtable E880 scores 3960, or 61.9 per core.
So, IBM won every benchmark, although Intel only suffered a minor defeat in per socket integer performance, where IBM's E870 was only 6% faster, by virtue of Intel's processor having 50% more cores. In floating point, even with this advantage, the E7-8890 lost by 39% in per socket performance. In per core performance, it goes from hideous to catastrophic. In SPECint*_rate_base2006, Intel's E7-8893 v2 loses by 37%, in SPECfp*_rate_base2006 by 48%. The peak performance favored IBM's processors considerably more, so I chose base to give Intel the best possible chance. Clearly, Intel is not quite at the performance level of IBM's server processors. More than that, the OpenPOWER platform is more capable in other ways as we will discuss hereunder.
As we go forward, please keep in mind, this is not a 2016 processor from AMD, with unknown capabilities, or a letter of intent from Qualcomm. It's here, and measurable. It also has an enormous amount of mature software available, unlike the ARM platform, which is only starting on this path. Furthermore, IBM has some very important partners in this endeavor.
On August 6th, 2013, IBM was joined by founding partners Google (NASDAQ:GOOG) (NASDAQ:GOOGL), NVIDIA (NASDAQ:NVDA), Mellanox (NASDAQ:MLNX) and Tyan, when it created OpenPOWER. Today, there are over 80 members, including newer Platinum members Samsung (OTC:SSNLF), Micron (NASDAQ:MU), Ubuntu, Altera (NASDAQ:ALTR), SkHynix and PowerCore. There are multiple purposes for this alliance. I think most readers already know IBM's server sales have been poor for the past few years, although they should recover somewhat with the release of the new z/13 mainframes. Even by IBM's own admission, the current markets it is in are flat, and it needs products for growth markets. Moving the POWER platform into the mid-range, and targeting Web 2.0, are primary goals of IBM.
To allow this, IBM has opened up the architecture, allowing partners to alter both the hardware and software stack. The Firmware, Operating System, and Hypervisor are all open source, giving partners the ability to customize the platform to best meet their needs. Integral to the platform's openness and performance is CAPI (Coherent Accelerator Processor Interface). This allows one to plug a co-processor into the system, and for it to have the same access to the caches and memory as the CPU. It completely bypasses the awkward hardware and software stacks inherent in I/O devices on PCI-E, for example, allowing for much higher throughput, and much lower latency. This is very exciting for companies like NVIDIA, where it can plug in a GPU compute module to boost highly parallel workloads, or for a company like Altera, where it can add an FPGA chip to assist with compression and decompression on the fly. Or Xilinx which will be contributing a FPGA memcached Key Value Store with excellent performance/power ratio, and very low latency. In fact, IBM and NVIDIA have already worked together to create the IBM Power Systems S824L server, combining POWER8, with NVIDIA's GPU computational technology, albeit through PCI-E, not CAPI. NVIDIA also will be contributing full CUDA support, and has co-developed NV Link with IBM, a high-speed processor interconnect, which will be used in future products from both companies.
One of the current trends among large server users, like Google, is the desire to have some input into how hardware is developed, so as to better tailor it to their situation. With OpenPOWER, this is now possible, and furthermore, they have already demonstrated a working system based on OpenPOWER, although it is currently being used only for testing, with no commitment to actual use. Even so, Google has said moving the software stack to POWER was easier than expected. It is also notable that Gordon MacKean of Google is the Chairman of the OpenPOWER foundation.
Other companies, like PowerCore have a license to build their own processors, which can extend the reach of the instruction set into other markets, while allowing IBM to produce the most powerful processors available. It is only one of several Chinese companies, with tacit support from the government, that have joined the OpenPOWER foundation. In fact, bowing to China's interest in smart cities, IBM and OpenPOWER partner Sichuan Huaxun Zhongxing Technologies Co. have opened a research center for developing smart city and Internet of Things technology in Sichuan province. Given China's sometimes hostile attitude towards American technical companies, this is significant even beyond the obvious technical implications.
Mellanox provides end-to-end interconnect technologies based on Infiniband, and Ethernet. Mellanox also contributed RDMA, or Remote Direct Memory Access technology. This allows very fast communications within a datacenter, with very low CPU utilization. Mellanox is also excited about utilizing their CloudX Platform on OpenPOWER. In their own words "Mellanox CloudX™ is a group of reference architectures that allows companies to build the most efficient, high performance and scalable clouds based on Mellanox's superior interconnect and off-the-shelf building blocks (servers, storage, interconnect, and software)."
Tyan is a highly-respected motherboard maker, and has already created a reference platform for OpenPOWER, with more to come. Officially, Samsung, SkHynix and Micron are "committed to supporting the OpenPOWER Foundation through the supply of memory and storage components for an open ecosystem." Whether or not this remains the extent of Samsung's involvement, remains to be seen, since they are a Platinum member, and that requires 10 full time employees to be dedicated to OpenPOWER.
While there are never guarantees for any technology, or any company, the OpenPOWER foundation appears to be a very viable alternative to the 'row-run' x86 boxes that command a large server market share. With a vastly more capable CPU, a more advanced and highly customizable platform, and openness that allows partner companies an extremely high level of innovation and configurability, this platform appears far more dangerous to Intel than AMD, Qualcomm, or any other ARM vendor. One has proven technology, with strong partners like IBM, NVIDIA, and Google, the others have yet to release products competitive with Intel's current x86 platform, yet often command more attention. Time may reveal them as strong competitors, but right now that is not very clear. This is not even considering the nod from China, and the desire of the Chinese government to make more processors there. Now they can.
There are still questions, mainly price. IBM's POWER platform is already very capable, yet it commands only a small (albeit, high-end) section of the market. Will OpenPOWER's consortium be able to bring the price into a range where the products can reach into a higher volume market? Will the openness of the platform continue to spur growth, and will companies in it create winning platforms based on the contributed technology, or will companies find the RoI in OpenPOWER isn't high enough to continue funding it after some point? Which brings up the bigger question of just how committed members, besides IBM, are. Will Google, actually use this platform, or will it never get past testing? Will the POWER architecture scale well to smaller parts, without extensive investment?
My personal belief is probably obvious, and I believe this has an excellent chance of success, and is the greatest threat to Intel and x86 in the server space. The technology is excellent, the degree of freedom partners have extraordinary and the partners strong and well-respected. We should also consider the importance of China, and IBM's partnerships there. While there's no such thing as a sure thing, this comes pretty close for me; the argument for it is compelling on so many levels.
Disclosure: The author is long AMD.
The author wrote this article themselves, and it expresses their own opinions. The author is not receiving compensation for it (other than from Seeking Alpha). The author has no business relationship with any company whose stock is mentioned in this article.