23 Apr IBM’s OpenPOWER and POWER8: Too Little Too Late?
IBM has long held the leadership spot in what the industry calls “engineered systems”, or big iron for large enterprises. IBM invented the mainframe and are very proud of that and the platform’s history in solving major business and eco-science problems. IBM’s Watson is huge part of IBM’s future and is a giant risk, but if IBM can pull this one off by attaching high margin revenue, it could again find itself in a leadership position that will make it difficult to be copied. But what about the future of private and public clouds where all the hardware growth is? IBM held a major announcement today in San Francisco which gave some insights into this future, specifically new IBM POWER8 systems and more details on the OpenPOWER Foundation. While this is an overview, we have published a white paper here (free). Let me start with a bit of background.
Before the Linux operating system became popularized, UNIX was king and each systems vendor had their own version of UNIX. Examples are IBM’s AIX and HP’s HP-UX. While more “open” than 100% proprietary operating systems like IBM’s z/OS, they were closed in the sense that they only ran on specific hardware from that vendor and feature support was many generations behind the UNIX kernel. HP and IBM both had custom processors, too, based on POWER, Itanium, and PA-RISC. IBM dominated this space peaking at 55% market share, but that market that has essentially cratered in size to roughly 10% of overall server revenues, the remaining 90% dominated by Linux and Windows servers.
Most of the percentage growth in the Linux server market come from these massive scale-out datacenters of the likes of Google, Facebook, Amazon, MSN, and eBay. These are stealthy, football field size datacenters with very fancy electrical sources and HVAC, many of them operating as a “lights-out” facility with limited staff. In some of these datacenters, if a rack fails, they just leave it for a week, pull it and may throw it away. Some are modular datacenters that are pre-assembled and flown in by planes and hoisted on top of buildings. This is a far cry from IBM’s traditional systems customers, and, quite frankly, IBM didn’t have much of a play there except for lower margin x86-based hardware that they are now selling to Lenovo. IBM needed a way to play in this scale-out market where even Dell and HP were having an increasing challenge to drive good, high-margin business.
While all the percentage growth was happening in scale-out datacenters driven by Intel’s Xeon, IBM was sitting on a systems and processor architecture that was very high performance. AMD’s departure from the high-end x86 server market left Intel with 100% market share. While ARM has licensed server capabilities to many, none of them are on track to hit Xeon E7 and high-end E5 capabilities in the near future, but rather lower end E5 and E3’s. This is where OpenPOWER comes into play.
So IBM has “opened up” the POWER platform and formed the OpenPOWER Foundation led by Altera, Google, IBM, Mellanox, Micron, Nvidia, Samsung, PowerCore and Tyan with a focus on scale-out hardware and software capabilities. This is a huge move and one very different for an IBM who has arguably focused on proprietary mainframes and “scale-up” technologies and systems.
OpenPOWER hopes to create an ecosystem similar to ARM’s in the mobile world with thousands of non-IBM participants involved in:
- fab technology (0 non-IBM members)
- SoC and IP development (5 members including Nvidia, Altera, Suzhou PowerCore, Xilinx, and VeriSilicon)
- I/O like memory, networking and storage (6 members including Mellanox, Fusion-io, Micron, Samsung, SK Hynix, Emulex)
- ODM and OEM systems (5 members including Tyan, Chuanghe Telco Tech, Servergy, Inspur, ZTE)
- firmware (1 member, Google)
- open source and application software (3 members including Teamsun, Google, Juelich).
It took ARM a decade to dominate in mobility and there is no reason that OpenPOWER would be any quicker. Except….. there’s the Google factor. If Google committed to move let’s say, “50% of their overall production workload to POWER”, things could move a lot quicker. This would get everyone’s attention, but keep in mind both Google and Facebook love to get on stage with anyone who could give them negotiating leverage over Intel, who has 100% of this business.
While I have a fair amount of “show-me” related to OpenPOWER, I have to say I love their approach to heterogeneous computing using accelerators. I have long held the belief that the scale-out datacenter of the future would be “workload and application specific”, not a homogeneous environment where all servers run the same, virtualized software. The scale-out future is where racks of servers are dedicated to different workloads using more specialty hardware that’s optimized for the jobs. IBM and OpenPOWER have created an environment where specialty accelerators from Nvidia, Altera, and VeriSilicon can be used to accelerate workloads far beyond what can be done today, and done in an architecture that’s easier to program than previous architectures.
To address heterogeneous computing, IBM created “CAPI” (Coherence Accelrated Processor Interface) which makes these accelerators first-class citizens to the processor and memory, not second-class citizens like all other architectures out there. So why does this matter? It means that programming to these accelerators is easier and you can get much, much higher performance, which if accepted as the future, could give IBM a leg up. To be fair, there’s nothing about CAPI that couldn’t be replicated by an Intel or the HSAF, but the fact that IBM and OpenPOWER has it now and partners are adding gates to their silicon and/or boards says a lot.
So what does all this mean to IBM? I believe POWER8 and associated IBM POWER8 Systems will scream running IBM software. POWER8 will make a great upgrade for those IBM customers running POWER7 who want to stay with the POWER architecture. Unfortunately, there isn’t any third-party data to suggest they will outperform Intel’s latest Xeon E7 v2 processors on like, non-accelerated workloads running the same software. IBM showed some very impressive numbers versus Intel’s Xeon E5 and very impressive figures using accelerators, but what will matter most is how they do versus the new E7 running the same software.
Net-net, OpenPOWER and POWER8 bring some very compelling technology and a good upgrade path for current IBM POWER customers. What’s unclear is how it will do in the massively growing scale-out markets, but time will tell as we see more third-part, independent benchmarks. Given OpenPOWER is in the “crawl” stage, we will also undoubtedly see some new things from Intel who dominates in this market. Who says big iron is boring?
Again, this is an overview, and if you want to see our deep dive on the subject, you can find a paper (free) here.