When Facebook open sourced the hardware and datacenter designs of its very first homegrown datacenter in Prineville, Oregon nearly eight years ago, creating the Open Compute Project, it was an act of enlightened self-interest. Facebook was growing fast enough that it could have, like peers Google and Amazon, kept its designs to itself and engaged with original design manufacturers (ODMs) to make the iron it specified, occasionally bragging to the world about the iron it had forged.
But instead, Facebook wanted to minimize its risk by cultivating a broader ecosystem of Open Compute system users, which in turn would compel a broader base of manufacturers to make systems that adhered to the standards it had created, many of which bucked the form factors and habits of long-established datacenter operators. By doing this, Facebook could create a virtuous cycle of innovation that it could participate in but not be entirely responsible for maintaining. The Open Compute effort, we think, has been successful, even if it has taken a lot longer to transform the ecosystem than many had thought.
The Next Platform was on hand at the OCP Global Summit last week in San Jose, and we are still chewing through many of the things we saw and heard there. We covered the innovative “Zion” hybrid CPU-GPU system created by Facebook to run its machine learning software already, but there was a lot more than this going on and we will get you the important bits over a series of articles. In the meantime, we wanted to talk about the Open Compute ecosystem itself and how it is taking root outside of the major hyperscalers who have contributed platforms alongside Facebook over the past several years and also talk about what is going to happen in the datacenter as many of the hyperscalers in the United States and China embrace the standards, such as the OCP Accelerator Module, that is embodied in the Zion machine.
The good news for those who don’t want to be strictly under the control of the dominant chip suppliers and the traditional OEMs who have been their main route to market for decades – and importantly for those who cannot bring hyperscale weight to bear like Google, Amazon, Microsoft, and Facebook in the United States and Alibaba, Baidu, Tencent, and China Mobile or JD.com in China can – is that the Open Compute ecosystem is continuing to grow in both size and coverage.
Mark Roenigk, who used to run the supply chain at PC and server maker Compaq way back in the day and then went on to run the hardware supply chain at Microsoft when it got into game consoles and did infrastructure stints at Intuit, eBay, and Rackspace Hosting, and who now is head of hardware engineering at Facebook, opened up the summit by explaining how far open hardware has come in the past decade.
“The theme of this conference in open together, and I think everybody in this audience will agree that we are better together,” explained Roenigk. “Back in 2010, when we started talking about an open technology forum from a hardware and datacenter perspective, a lot of companies did not share information and their hardware and datacenter designs were proprietary. Look at what has happened in the past nine year. This community has been a great success, but of course the community is only as good as the commitment and passion of that community.”
There are now 178 member companies in the Open Compute effort, with 20 of them added in the past year, and they collectively have thousands of engineers working on projects that feed from and feed back into that community. And it is now just about hardware. There are two software companies that have joined, and given how integral systems software – BIOS, other firmware, management systems, middleware, and so on – is to any hardware platform, we expect for the number to grow, and we might even go so far as to say that the Open Compute Project should probably give some consideration to the idea of merging with the Open Infrastructure software community – formerly known as OpenStack, the open source cloud controller started by Rackspace and NASA six months before Facebook launched Open Compute. But that is a story for another day. At this year’s OCP Global Summit, there were 124 new products being rolled out, 27 percent more than at last year’s summit, and around 60 percent of these products will be available within the next two months and of these, the majority of the ones shipping fastest are in networking.
“If you had told me back in 2011 when we started OCP that we would be a major force in networking, I would not have thought that this was even possible,” said Roenigk.
The success of Open Compute in networking, we think, is more a reflection of how locked down and proprietary networking was before the idea of prying the boxes open and releasing the software for faster – and less proprietary – innovation became the mantra of the hyperscalers and big public cloud builders of the world. (It is not a coincidence that Facebook’s head of networking, Najam Ahmad, has perhaps been the most vocal about smashing open the proprietary switch.)
The point is, 2018 has a much more voluminous product catalog than in years gone by, and products are coming to market more quickly, too.
“Open Compute has done a really good job in driving efficiency, which means getting the right engineering talent around the table and to decide what the next generations of technologies coming down the pike are and how we can develop standards around those and commonalities around different industries,” Roenigk continued. “We also continue to focus on efficiency, and we have developed a power shelf that utilizes 98 percent of the power in a particular rack, and you people know that the industry standard just a few years ago was 80 percent. We have higher performing compute and we are using a lot less power, and now that we are coming up against some basic laws of physics, we have a ton of new challenges ahead of us. So I think our work here at OCP is literally just getting started. We also need to address the needs and demands of the industries that we support. Many people think that Open Compute is just a technology forum, and while it is about infrastructure, we serve oil and gas, financial services, and certainly the tech industries, and increasingly we work on data analytics and artificial intelligence.”
One of the key benefits of Open Compute is that members have to contribute – they can’t just “pay to play,” as Roenigk put it. Member companies working in specific areas have to contribute, and therefore open source, intellectual property. Small wonder, then, that hear so little from Apple, which joined OCP four years ago, didn’t do much or say anything, and is no longer on the official member list as far as we can see. Amazon and its cloud division, Amazon Web Services, never have joined, but its infrastructure guru, James Hamilton, did speak at a conference many years ago without ever joining. Amazon does not appear to be a joiner or a sharer. Microsoft pulled a gambit maneuver and open sourced its internally designed systems in 2015, creating a whole second stream of open hardware alongside that of Facebook. Google joined three years ago to work on converged rack designs mixing the benefits of traditional 19-inch racks and 21-inch Open Rack enclosures, and is in particular pushing 48 volt racks. Alibaba, Baidu, Tencent, and JD.com are all top-end, platinum-level sponsors, and Rackspace is still in there but like Goldman Sachs, and Intel are not really hyperscale. AT&T and Deutsche Telekom, also OCP members, have the potential to be.
Collectively, these user companies and those that emulate them can bring a lot of weight to bear – a lot more than Facebook could ever have done by itself. Trying to figure out how much weight that is, well, that is a tricky business indeed. But it is getting easier in some respects.
Last year, the Open Compute Foundation commissioned IHS Markit, one of the big market researchers that specializes in tracking chips and infrastructure sold around the world, to do a comprehensive study of the size of the market for OCP servers, storage, networking, racks, power shelves and distribution, peripherals, and related stuff. The results of that 2017 market study were released in February 2018, in case you missed it, and Cliff Grossner, executive director of research and analysis, and Vlad Galabov, principal analyst, at the company unveiled the results of the 2018 study of the OCP ecosystem at the Global Summit last week, which cases the OCP market outside of the sales to the OCP board members in 2018 and updating its forecasts going out to 2022.
That distinction between the entire OCP market and the sales of OCP board members is significant and annoying, and you can blame the OCP board – comprised of executives from Facebook, Rackspace, Microsoft, Intel, and Goldman Sachs – for not being, well, perfectly open about it. To be fair, there is a logic to looking at the health and expanse of the OCP ecosystem in terms of numbers of customers, revenues, shipments, industries, and geographies outside of the key companies – most importantly Facebook and Microsoft, which are probably the dominant buyers of OCP gear in the world, and Microsoft probably tops Facebook at this point – to see how it is spreading to the IT market in general. But by cutting out any information about the board members that also radically underrepresents the market penetration in the worldwide IT market, and this is silly. It’s just funny how the companies that are so eager to gather and sell or otherwise profit from our information are so secretive about their own. OCP could have – and should have – provided the complete set of information, without giving market shares individually for the OCP board members. That was all that was required to protect their information.
We can still learn a lot of about how the OCP ecosystem is doing outside of shipments to Facebook, Rackspace, Microsoft, Intel, and Goldman Sachs. In 2017, sales of OCP gear to the community aside from that of the OCP board grew by 103.1 percent to $1.19 billion, according to IHS Markit, with servers accounting for nearly three quarters of those sales. To be precise, non-board OCP server sales in 2016 were $373.6 million, and rose by 130 percent to $859.4 million in 2017; switch sales rose by 81.2 percent to $102.6 million over the same period, and storage rose by 32 percent to $161.9 million. Toss in the other stuff, and you get to that $1.19 billion figure for 2017. By the way, that gave the non-board OCP buyers a 0.87 percent share of the total addressable market, which Grossner pegged at $137 billion in 2017.
For the interim 2018 report released at the OCP Global Summit last week, IHS Markit is taking the first three quarters of sales of non-board OCP buyers and adding to it estimates from the fourth quarter; the final four quarters of results will not be done for a few weeks. Moreover, we have not been able to get the precise drilldown on revenues by product, but Grossner did flash up this chart to give us an idea:
Galabov said that sales of OCP servers across the non-board buyers was a lot higher than expected in 2018, in part due to GPU, GPU, memory, and flash prices being higher than expected and also because companies are buying richer configurations. Sales might have been higher if some companies did not delay their projects by a quarter or two, according to Galabov. In any event, as you can see, server sales kissed $2 billion not including those secretive board members, somewhere around a 2.25X factor of growth, and it looks like it will grow around 50 percent this year and another 40 percent in 2020, reaching up to around $7.5 billion by 2022. This is a substantial part of the server racket, and when you consider that the board members might have – even then – two or three times as many OCP server purchases, it looks like OCP could turn out to be a category unto itself when it comes to server infrastructure. OCP will be larger than any other kind of product line, with a mix of ODMs and OEMs making the gear, all to open standards. Retrofitting Open Compute gear to fit into standard 19-inch racks seems to be a key driver of growth, since enterprises do not want to retool their datacenters for Open Racks – even if Facebook is right about everything when it comes to criticizing the 19-inch rack standard.
The push to the edge might see Open Compute iron adopted at a much higher rate in remote facilities, 5G base stations, and other service provider edge sites. The TAM is going to expand, and companies are not necessarily going to want to get locked into proprietary gear out there on the edge. Everyone has learned that lesson, for sure.
There are a lot of reasons why companies are adopting OCP iron, and as you might expect, cost reduction and power efficiency are the big drivers. But having flexibility across suppliers and designs is also a key factor:
But there are still a lot of barriers to adoption as well, and they are the same as with adopting any new technology:
While the ODMs do a fairly good job of fulfilling demand, enterprises are more comfortable with their established and well-known OEMs, some of whom support OCP equipment, some of whom do it begrudgingly, and some of whom would rather not.
There is an interesting switch underway in the OCP customer base, and that is a move away from direct sales with hardware exchange when something fails to acquisitions through system integrators on a rack scale, according to Grossner. Here is what the situation looks like now among the OCP gear buyers who are not members of the OCP board:
Some of this shift from direct to system integrator acquisitions is related to the ramp of OCP iron among telecommunications companies and service providers, who historically like to work through system integrators that deliver at the rack scale rather than the ODMs that were raised in stature substantially by the hyperscalers and cloud builders. The system integrators often provide the support and warranty services for the iron they push, and they want someone with global reach to provide that. In essence, the system integrators fill the gap between OEM and ODM.