Input sound file:
Session Name: E Future Of Data Center Servers And Processors: The Death of One-Size-Fits-All In The Server World
Announcer Andrew Feldman
Hi there, everyone. We’re ahead of schedule, so why don’t you turn your cellphones back on and tweet or something, waste time so we’re back behind schedule. No, let’s not do that. Instead, we have a really cool presentation coming up on the future of data centers, and I was talking backstage with our next speaker, Andrew Feldman, and it’s a pretty interesting talk about what’s going on and what we can expect next from the cloud. If there’s one thing that tends to be true about Structured, it’s that you hear about the latest ideas here first. This next talk and all other talks hopefully, will present you with some really interesting ideas. So without further ado, please join me in welcoming Andrew, who just got back from Florence Italy, so hopefully he’s not too jetlagged. Andrew, a pleasure.
Andrew Feldman 00:59
Good morning, everybody. We’re going to talk a little bit about the drivers behind fundamental changes in infrastructure. We’re going to do it in a very unusual way. Usually when we talk about infrastructure, we begin with infrastructure. We’re not going to begin with infrastructure, we’re going to begin with legal stuff [chuckles]. We’re not going to do that. What we’re going to begin with are two photographs. They’re interesting photographs, particularly for someone who’s Jewish to present. These are two photos from the exact same vantage point at the Papal inauguration in 2005 and 2013. What they show is a complete transformation of the client side of the world. Not only do they show a change on the client side, they show a complete change in the way we interact with the world. Remember, this isn’t Silicon Valley, what this is every single person watching an important event, creating video of it, on a client side device, and then pushing that client side information up into the cloud, either on purpose or through their backup, or by sending an email to their aunt or uncle or posting it on Facebook and Twitter.
Andrew Feldman 02:22
I could find no photograph that more thoroughly captured how the demand for compute has left the client side and moved into the data center, and here’s why. Everybody have their phone with them? Everybody? Could you hold for a sec? Now turn off data services. What I’ve just done is I’ve taken the cloud away from you. I’ve turned your phone into an Angry Birds player.
Andrew Feldman 02:58
That’s right, it is now an absolutely uninteresting phone that plays Angry Birds. That’s the point. The point is what is interesting, what you want to do with that device, is get to the cloud. What is even more interesting is that the devices in your hands for the most part do no interesting work. What they are very good at though is displaying the work done elsewhere. That change is profound and multifaceted. How many of you here today are working on an iPad? I see at least one here. Do you know what processor it has in it? Do you care? Right. That’s a really interesting thought as well, you don’t care. The experience of working with it interesting, it’s different. It’s more valuable to you than putting a much faster processor down. When you look at the new client side of the data center, it is doing no work. It is pushing the demand for compute into the data center, and in the data center we enable all your client side devices to do anything you want.
Andrew Feldman 04:30
I will in this slide explain to you why software defined networking, the change of storage, and the change in compute is all happening, because we have more client-side devices, we have more apps, we have more users. They are driving change upstream in the data center. So profound is this change, it’s not just compute, it’s not just networking, and it’s not just storage that’s changing, but the very buildings we house them in are changing as well. The locations we put the buildings are changing. We used to put data centers near urban environments. Where do we put them now? We put them in eastern Washington, along the Deschutes river in Oregon, to get low cost power. They used to be regular tilt slab buildings, not very different than a Costco, and now they’re extremely engineered facilities, to use ambient air 330 days a year to drive down the cost of air conditioning, to do a hundred things. It’s not just compute and storage and networking that’s changed, but even the buildings that house them have undergone profound transformation. This isn’t going to stop.
Andrew Feldman 05:44
A third of the world’s population is connected to the internet. How are we going to get to the next two thirds? We’re going to get to the next two thirds by phone. There are six billion mobile phone subscriptions out there, and only a billion people with a phone that has data access. So we know how we’re going to get to the rest of the world. We’re going to get to them on their phones, and what we know is that their phone doesn’t have a ton of compute in it, and that compute is going to get delivered to the phone, but done in the data center. Now it’s easy in these discussions to point to the third world and say “Look, all this growth is coming out of Brazil, or India.” It is, but here’s an interesting analysis of the first world, in the US. This speaks directly to the duration of time we spend in the cloud. Over a three year period we went from basically zero to over a third of the adult US population having a tablet. What does that mean for us? We now take our tablet on vacation. Anybody sit and annotate television with their iPad? Ever watch a baseball game, see a good play and wonder how much he makes? This is now a way we spend hours and hours a day in the cloud, when before we were on the couch. So it’s not the third world, it’s devices, it’s users, it’s applications creating duration of time that we spend, and that’s driving such profound change in the data center for every aspect of the data center we build.
Andrew Feldman 07:38
What does this as we look forward? It means on the client side, it is not about CPU performance. As I look around and see all the tablets in the audience, it is proof of this claim. The client side is about your interaction with the device, it’s about experience, it’s about the graphics. We can learn this either by thinking really hard, or by looking at what Apple chooses to market. Have you ever heard Apple talk about which processor they’re using? What they talk about is the retina display. Why? Because that has the biggest impact on your experience. It’s the graphics, it’s the display and the visualization of your interaction, which is the driver going forward for client side devices. On the server side, this has created tremendous changes as well. The data center now does the compute for the client side. That means millions and millions of users, each with parallel work. That’s what we ask the data center to do, we don’t ask it to do CADCAM, we don’t ask it to design airplane wings. The vast majority of the work that it’s doing is simple parallelized work that was generated on the client side. This is by the way what Google and Facebook and Ebay and Amazon and Ali Baba, that is what they do. They generate work for servers, and that work is very different.
Andrew Feldman 09:41
That work being different provides and requires a different type of server, like the SeaMicro SM15000, or HP’s Moonshot, or Dell’s Viking, or any number of different machines, but not the same architecture. It also requires a different type of processor, a more efficient processor. Increasingly we believe it’s a single-socket processor, and in the future we believe it’s going to be an ARM processor. So when we look forward, what are the things that in the compute industry, and I’m sure later today you’ll hear from the storage industry, you’ll hear from people in the SDN world, and how they’re responding to the fundamental transformations that the client side has wrought on the data center. In our view, the transformations in the server world are in four dimensions, and we should demand them of ourselves, of our partners and of our vendors.
Andrew Feldman 10:52
The first is the death of the one-size-fits-all server CPU. What we do now is we match workload to processor type. That mean different types of processors. The vast majority of processors historically that did server workloads were two-socket processors, where performance was the single most interesting characteristic. Now it’s performance per watt per dollar which is the most interesting characteristic. Now single-socket CPUs are enormously interesting, not only single-socket CPUs but CPUs that include the graphics engine, called an APU, to do different types of work. They are a better fit, that’s one dimension. The second dimension is it’s time we embrace another instruction set. If you look at the result of a flourishing ecosystem, which is what you can see in your handheld devices, we at AMD are embracing ARM for servers. We think that’s a direction that the world will go. I think we also need to do more other interesting things. We need to invent more efficient servers, not just the CPUs. For a long time in our industry, we just sat around and we waited for the next more efficient processor to come out. We put that on a motherboard in the same way we’d made a motherboard for years and years. Those days are over. We need to invent more efficient servers. Finally, we need to embrace open-source hardware. Obviously the pioneers there have been Facebook and Frank Frankovsky and his team, but we ought to be building and pushing up designs so we all don’t do the same thing again and again.
Andrew Feldman 12:59
Now I’m going to leave you with a controversial thought. In the sixty year history of compute, smaller, higher volume parts have always won on the server side, without fail. I know many of you are in the finance world, and you guys have this disclaimer that past performance doesn’t [inaudible], right? In fact, in our business it does. Every single company that has fought this has lost. This is why at we at AMD look forward, and as I look forward as a strategist, and I think as many of you look forward, when you see a 20x advantage in volume in a volume dependent industry, it’s time to sit up and take notice. Manufacturing economies can’t overcome this. So with that, I will leave you with a thought that this type of history is rarely wrong. As you wander through today and the rest of– whether it’s a ball game, take a look at the way the client side devices are being used, ask yourself five, seven, ten years ago were they used the same way? Then ask yourself what changes were necessary, all through the chain up to the data center. Think about how that impacts every vendor who makes hardware, or servers, or CPUs, storage–
Andrew Feldman 14:45
Just one beat off. With that I’m done, thank you very much.