The launch of two new features into the Open Compute hardware specifications on Wednesday has managed to achieve what Facebook has been threatening to do since it began building its vanity-free hardware back in 2010. The company has blown up the server—reducing it to interchangeable components.
With this step, the social network has disrupted the hardware business—from the chips all the way up to the switches. It has also killed the server business, which brought in $55 billion in revenue for 2012, according to IDC estimates.
It’s something I said would happen the day Facebook introduced the Open Compute Project back in April 2011, and which we covered again at our Structure event last June when we talked to Frank Frankovsky, vice president in charge of hardware and design at Facebook.
What Facebook and others have done—in a relatively short time by hardware standards—is create a platform for innovation in hardware and for the data center that will allow companies to scale to the needs of the Internet, but also to do so in something closer to Web time. The company will do this in a way that creates less waste, more efficient computing, and more opportunity for innovation.
So what is Facebook doing? It’s contributed a consistent slot design for motherboards that will allow customers to use chips from any vendor. Until this point, if someone wanted to use AMD chips as opposed to Intel chips, they’d have to build a slightly different version of the server. With what Frankovsky called both “the equalizer” and the “group hug” slot, an IT user can pop in boards containing silicon from anyone. So far, companies including AMD, Intel, Calxeda, and Applied Micro Circuits are committing to building products that will support that design.
The other innovation that’s worth noting on the Open Compute standard is that Intel plans to launch a superfast networking connection based on fiber optics that will allow data to travel between the chips in a rack. The company is expected to announce this 100-gigabit Ethernet photonic connector later this year, and Frankovsky can’t wait to get it into production in Facebook’s data centers.
I’ve written a lot about how we need to get faster, fiber-based interconnects inside the data center, and about the efforts to do so. What’s significant here is not just that this design speeds up chip-to-chip communication. With the right hardware, it runs a rack into a server and makes the idea of a top-of-rack switch irrelevant—something that Cisco Systems, Juniper Networks, and Arista Networks might worry about (although Arista’s Andy Bechtolsheim is helping present this technology at Open Compute, so I imagine he has a plan). It’s also worth pointing out that only in 2012 did Facebook transition to an all-10-gigabit-Ethernet in its data centers, but now it wants to speed up the interconnects as soon as it can.
With its new features, Open Compute has managed to give customers—from financial-services firms to Web properties—a platform on which to build custom and modular servers.
A good example of this might be building a house. There are plenty of ways of doing it, from hiring an architect to choosing a plan from a selection offered to you by a developer. That was the former server world—you either went all custom or chose from “plans” provided by Dell and others. But those chosen plans might come with four bathrooms, and you might only want three. If you buy 10,000 unneeded bathrooms, that’s a lot of added cost that brings you no value.
With Open Compute, Facebook showed people how to build houses out of shipping containers. As more and more elements like Group Hug or standard interconnects are added, it’s as if you can pick your favorite brand of shipping container and pop it on. Even better, if you want a new bathroom, you can swap it out without ever affecting your bedroom. And for many companies, the cost of building and operating these new servers will be less than purchasing mass-produced boxes designed by others.
This is great for Web-based businesses and anyone who relies on IT. It cuts out waste, and it turns what once was a monolithic asset into something that can be upgraded with relative ease—and at a lower cost. So when Frankovsky talks about renting CPUs from Intel, for example, the Group Hug design makes that possible. He can rent CPUs for a workload and then use those servers for something else without chucking them. He can swap out chip architectures, networking cards, and drives at will to put the right mix of storage, network, and compute together for his jobs as they grow and shrink.
These servers can’t be provisioned as fast, but what Open Compute has done is create a server architecture that can scale and change as much as the cloud itself does. And it did so without the server guys.
Also from GigaOM:
Cloud and Data Fourth Quarter 2012 Analysis (subscription required)