 From Denver, Colorado, it's theCUBE. Covering Supercomputing 17, brought to you by Intel. Hey, welcome back everybody. Jeff Frick here with theCUBE. We're at Denver, Colorado at the Supercomputing conference 2017, about 12,000 people talking about really the outer edges of computing. It's pretty amazing. The keynote was these huge square kilometer array and new vocabulary word I learned today. So it's pretty exciting times and we're excited to have our next guest. He's Bill Jenkins. He's a product line manager for AI on FPGAs at Intel. Bill, welcome. Thank you very much for having me. Nice to meet you, nice to talk to you today. Absolutely, so you're right in the middle of this machine learning AI storm which we keep hearing more and more about. Kind of the next generation of big data, if you will. That's right, that's right. It's the most dynamic industry I've seen since the telecom industry back in the 90s. It's just evolving every day, every month. So Intel is making some announcements, using kind of this combination of software programming and FPGAs on the acceleration stack to get more performance out of the data center. I get that right? Sure, yeah, yeah. Pretty exciting the use of both the hardware as well as software on top of it to open up the solution stack, open up the ecosystem. So what are those things are you working on specifically? So I really build the first enabling technology that brings the FPGA into that Intel ecosystem where Intel is trying to provide that solution from top to bottom to deliver AI products into that market. FPGAs are a key piece of that because we provide a different way to accelerate those machine learning and AI workloads. Where we can be an offload engine to a CPU. We can be in line analytics to offload the system and get higher performance that way. And we tie into that overall Intel ecosystem of tools and products. Right, so that's pretty interesting piece because the real time streaming data is all the rage now, right? Not in batch, you want to get it now. So how do you get it in? How do you get it into the database? How do you get it into the microprocessor? So that's a really, really important piece. That's different than even what two years ago we really didn't hear about much real time. Yeah, I mean, I think it's like I said, it's evolving quite a bit. Now a lot of people deal with training. It's the science behind it. The data scientists work to figure out what topologies they want to deploy and how they want to deploy them. But now people are building products around it. And once they start deploying these technologies into products, they realize that they don't want to compensate for limitations in hardware. They want to work around them. So a lot of this evolution that we're building is to try to find ways to more efficiently do that compute and what we call inferencing, the actual deployed machine learning scoring, as they will, in a product. It's all about how quickly can I get the data out? It's not about waiting two seconds to start the processing in an autonomous driven car where someone's crossing the road, I'm not waiting two seconds to figure out what's a person, I need it right away. So I need to be able to do that with video feeds, right off a disk drive from the Ethernet data coming in. I want to do that directly in line so that my processor can do what it's good at and we offload that processor to get better system performance. And then on the machine learning specifically, because that is all the rage and it is learning. So there is a real time aspect to it. You talked about autonomous vehicles, but there's also a continuous learning over time that's not necessarily dependent on learning immediately, but continuous improvement over time. So what are some of the unique challenges in machine learning and what are some of the ways you guys are trying to address those? Yeah, I mean, once you've trained the network, people always have to go back and retrain. They say, okay, I've got a good accuracy but I want better performance. So then they start lowering the precision and they say, well, today we're at 32-bit, maybe 16-bit, and then they start looking at into eight. But the problem is their accuracy drops. So they retrain that into eight topology, that network to get the performance benefit but with the higher accuracy. But the flexibility of the FPGA actually allows people to take that network with the 32-bit train weights but deploy it in lower precision. So we can abstract away the fact that the hardware is so flexible, we can do what we call floating point, 11-bit floating point, or even 8-bit floating point. Even here today at the show, we've got a binary and ternary demo showcasing the flexibility that the FPGA can provide today with that building lock piece of hardware that the FPGA can be and really provide not only the topologies that people are trying to build today but tomorrow, future-proofing their hardware, but then the precisions that they may want to do. So they don't have to retrain. They can get less than a 1% accuracy loss but they can lower that precision to get all the performance benefits of that, data science's work to come up with a new architecture. But it's interesting because there's trade-offs, right? There's no optimum solution. It's optimum as to what you're trying to optimize for. So really the ability to change, the ability to continue to work on those learning algorithms to be able to change your priority is pretty key. Yeah, I mean, a lot of times today you want this. So this has been the mantra of the FPGA for 30 plus years. You deploy it today, it works fine. Maybe you build an ASIC out of it, but what you want tomorrow is going to be different. So maybe if it's changing so rapidly, you build the ASIC because there's runway to that. But if there isn't, you may just say, I have the FPGA, I can just reprogram it to do what's the next architecture, the next methodology. So it gives you that future-proofing, that capability to sustain different topologies, different architectures, different precisions, to kind of keep people going with the same piece of hardware without having to, say, spin up a new ASIC every year, which is, even then, it's so dynamic, it's probably faster than every year the way things are going today. So the other thing you mentioned is topography, and it's not the same topography you mentioned, but this whole idea of edge, right? So moving more and more compute and store and smarts to the edge, because there's just not going to be time, you mentioned autonomous vehicles, a lot of applications to get everything back up into the cloud, back into the data center. So you guys are pushing this technology not only in the data center, but progressively closer and closer to the edge. Absolutely, I mean, the data center has a need. It's always going to be there, but they're getting big, right? There's the amount of data that we're trying to process every day is growing, right? I always say that the telecom industry started the information age. Well, the information age has done a great job of collecting a lot of data. We have to process that. If you think about where, again, I'll maybe I'll allude back to autonomous vehicles, you're talking about thousands of gigabytes per day of data generated, smart factories, exabytes of data generated a day. What are you going to do with all that? It has to be processed. So we need that compute in the data center, but we have to start pushing it out into the edge where I start thinking, what are you going to show like this? I want security. So I want to do real time weapons detection, right? Security prevention. I want to do smart city applications, just monitoring how traffic moves through a mall so that I can control lighting and heating. All of these things at the edge in the camera that's deployed on the street, in the camera that's deployed in a mall, right? All of that, we want to make those smarter so that we can do more compute to offload the amount of data that needs to be sent back to the data center as much as possible. Relevant data gets sent back. No shortage of demand for compute store networking, is there? No, no, it's really a heterogeneous world, right? We need all the different compute. We need all the different aspects of transmission of the data with 5G. We need disk space to store it. We need cooling to cool it. It's really becoming a heterogeneous world. All right, I'm going to give you the last word. So I can't believe what we're November of 2017, which is bananas. What are you working on for 2018? What are some of your priorities? If we talk a year from now, what are we going to be talking about? Yeah, so Intel, I mean, Intel's acquired a lot of companies over the past couple of years now on AI. You're seeing a lot of merging of the FPGA into that ecosystem. We've got the Nirvana, we've got Movidius, we've got Mobileye acquisitions, Saffron technologies, all of these things when the FPGA is kind of a key piece of that because it gives you that flexibility of the hardware to extend those pieces. You're going to see a lot more stuff in the cloud, a lot more stuff with partners next year and really enabling that edge to data center compute with things like binary neural networks, ternary neural networks, all the different next generation of topologies to kind of keep that leading edge flexibility that the FPGA can provide for people's products tomorrow. Exciting times. Yeah, great. All right, Bill Jenkins. There's a lot going on in computing. If you're not getting computer science degree kids, think about it again. He's Bill Jenkins, I'm Jeff Frank. You're watching theCUBE from Supercomputing 2017. Thanks for watching. Thank you.