 From San Jose, in the heart of Silicon Valley, extracting the signal from the noise. It's theCUBE, covering OCP US Summit 2016, brought to you by OCP. Now your host, Jeff Frick and Stu Miniman. Hey, welcome back everybody. Jeff Frick here with theCUBE. We are live in San Jose at the Open Compute Project Summit 2016 OCP Summit. This is our third year bringing theCUBE here. It's really an interesting story because this is the cloud and really it was initially a Facebook initiative to open source a lot of their hardware infrastructure to share with the world how a hyperscale company is developing this stuff and then share the secrets, kind of open the kimono. And now all the hardware vendors are jumping on board to support this standard. So now we're actually going to talk to a hardware guy who's stuck in a software company but we're happy to have him on Kevin Lee, technical program manager from Facebook. Welcome, Kevin. Thank you. So you already had a session today. You got a session tomorrow. Your session's all about Big Sur. Tell the folks in the audience, what is Big Sur? So Big Sur is our purposely built hardware for artificial intelligence. So it's intended to do a large model deep learning applications. So what are some of the things that make it specific to that application that you had to really take into consideration? Yeah, so it was designed around NVIDIA's M40 GPUs. So it houses eight GPUs up to 300 Watts each. So compared to our other OEM solutions that we've used in the past, the speed is twice as fast and we can train models twice as large with this system. So are there any examples you can share of some of the applications where this thing is being applied? Yeah, so you can see it all on Facebook already, the type of AI and machine learning algorithms that we do train on it. Can you give me one? Yeah, so. Or what process maybe is better, is improved by the application of Big Sur that I wouldn't notice or maybe I wouldn't notice it just magically happens for me. So as Jay mentioned in his keynote today, there's a lot of applications that are used in ads or used in facial recognition and stuff like that. So that's what it's used for. Okay. So Kevin, can you give us a little bit of insight is kind of the relationship between kind of the hardware and software folks inside of Facebook. I think back to when OCP first launched and we said, boy, companies like Facebook, Google, Yahoo, have these team of PhDs, they really build an application and infrastructure just serves that but it's very different from what you saw in the enterprise. Maybe give us your experiences and how the teams work and what might be different from what most companies say. Okay, yeah, so for Big Sur specifically, when we designed and built Big Sur, it was targeted at a certain application. So a lot of the models and training that we do are very specialized for AI and a lot of the offerings that the OEMs provided at that time didn't have the sort of architectures that we wanted. So Big Sur in itself was actually designed together with the software in mind. And then how does it feel to kind of let your baby out? You know, right behind you is the Facebook booth. There's been a massive line there all day standing around that big rack and there's a Big Sur server in there, correct? So how does it feel to kind of open it up and share it with the world outside of the walls of Facebook? I mean, it feels amazing. So we spent about 16 months working on this hardware. A lot of literal blood and sweat went into this design. Hopefully not too much blood, not too much blood. So we went into making it purposely built for AI and that's what we did, yeah. All right, so Kevin, you know, if you've had some conversation in the show, what kind of questions are you getting from attendees here? Give us a little bit of vibe of what you're getting. So some questions that were asked were how do I manage different type of PCIe topologies? How does it fit to AI's application specifically? So when we designed the Big Sur system, we made it so that it is flexible and dynamic. It has different building blocks, as we call it, inside the system itself and then we can actually put a standard motherboard into the system and different types of GPUs such as NVIDIA, AMD, Intel, they can all fit into the system. And you said you open sourced it in December. So what type of contributions are you getting back from the community that you guys hadn't thought of yourself? Yeah, so basically back to the software point too. So we've been actively contributing a lot of the software developments that we've done from the Facebook AI research team and being able to contribute the hardware is actually one big leap in this whole space for open source. So obviously Facebook started everything going on in OCP. What kind of direction do you get as an organization for kind of contribution, participating not only in your project, but what goes on in GitHub? Once something's out the door. Can you repeat that? So talking about just in general, your participation not just building initially but participating some of the community activity. Yeah, exactly. So the intent of Big Sur wasn't to throw it out there and then this is what you must use. So Big Sur itself, we are very open to getting feedback on the system itself. I'm sure there are faults in the system and we would iterate and build on that for the next generation. So OCP is very, very good at doing that. So you stand on the Big Sur project or are you off the new pastures? So currently I work as a technical program manager. I work both on the AI platform stuff and also on our signal socket architecture. So we already talked about it today in the keynote of the Yosemite platform. So do you see all this horsepower bringing, coming to bear and you're the one that's actually putting it together and seeing the application and the performance? What gets you excited about what's going to be, that you'll be able to do six months, nine months, 12 months, I don't even say years, that's way too far in this scale. So back to Facebook's philosophy, what we're trying to do is connect the next billion people. So a lot of the work that goes into AI is to understand the world a little better. So I think Big Sur really connects back to our mission as a company. So Kevin, one of the things we've looked at with AI specifically is, how do I get an order of magnitude better, faster processing? So I guess what has been the biggest leaps forward that you've been talking about, kind of the density, is it the way you can redesign the airflow and where do you see the opportunity that if you look down the road and you say, oh, I can't wait to do a next generation of this because boy, it's going to be so much better, faster, cheaper a year from now. So the field of AI is very interesting at this moment. So there are many different offerings that is out there. So to understand what works well in our infrastructure, that's the very exciting part that we have today. So currently, Big Sur right now, it supports up to AGPUs so models can fit up to the amount of memory that's on those GPUs. So in the future, maybe the models might get a little bigger and we have to think about how to actually extrapolate that into the rack or the data center level. So that's the part that gets very, very interesting. All right, Kevin, well before we let you go, give a little plug for your session tomorrow. I know you had more of a keynote today. Tomorrow's going to be a little bit more in depth hands on. Tell the folks what to expect tomorrow and when and where they should show up. So tomorrow I'm doing an engineering workshop, basically on the same topic that I talked about today. I'm going to go in depth about the PCIe topologies. It's at 10.30 tomorrow and anyone can come in and ask me any questions they'd like. Awesome, well Kevin, thanks for spending a few minutes. Kevin Lee, technical program manager from Facebook. These are the guys that basically do the hard work behind the scenes that make the magic happen where you can talk to your phone and ask it where to go or ask it a question. Really exciting stuff. I think of Mars Law, never talked about enough. Where this stuff is going, how fast it's going is really amazing. So Kevin, thanks for stopping by. Thank you. I'm Jeff Frick with Stu Miniman. We are live in San Jose at the OpenQQ Project Summit. 2016, we'll be back after this short break with our next guest. Thanks for watching.