 Live from New York, it's the Cube, covering Micron Summit 2017, brought to you by Micron. Welcome back to New York City, everybody. This is the Cube, the leader in live tech coverage. And we've got wall-to-wall coverage of Micron Summit. We saw announcements of NVME, NVME of a fabric, some great presentations from industry analysts like Laura Dubois, some practitioners, CIOs, and so forth. Steve Pausky is here, he's the Vice President of Advanced Computing Solutions, at Micron, he's joined by Tom Ebe, who's the Vice President of Compute and Networking Business at Micron. Gents, welcome to the Cube. Thank you. Great to be here, thanks. Steve, you were presenting today, and I got to start off. So I've asked this question of Pat Gelsinger, and he got very defensive as a former Intel guy, but so I'll ask you too. We said Pat is Moore's law attenuating, and he said, I won't tell you what Pat's answer was, but I got to ask you. I mean, we see it in the comments that you guys make to Wall Street, but is it really? How should we think about Moore's law? Well, I probably won't get as defensive, but when you spend 32 years defending something, it's hard to make a comment to the opposite. I think the most important thing is it's really, there are two elements that there's Moore's law, which is just all about doubling the number of transistors, and even with 16 nanometer, 14 nanometer, 10, seven, there's still that doubling is gonna happen for the next two or three process generation, so the doubling will continue. The big issue is Denard scaling. So Denard scaling was after, I believe he was an IBM fellow, and he observed that as you scale in terms of transistors because there was such an impact in terms of the turn-on voltage of the transistor to the supply that you saw benefit in speed and a benefit in power. In about 2006, Denard scaling started, we started to see that that wasn't happening. We were hitting a power wall and a power limit. So I wouldn't say Moore's laws come to an end, but Denard scaling has come to an end, so it's causing us to rethink how we build our systems because we just can't keep up with the performance requirements. And to Pat's point, Pat's like the greatest all-time. I used to work for Pat, so I know exactly. He's the greatest all-time Q-guest, and number one. Really, I'm hoping to unseat him. Okay, you need about 15, 20 more interviews. He's a, he's a, he's a machine. Quantity, not quality, okay. Oh, it's both. And he has the longest interview. John Furrier, my business partner and co-host, he could talk a lot, and at one point he just said, Pat, I give up, he talked to us. But his point was that we've marched to the cadence of Moore's law in terms of where the innovation has come from in this industry for decades. But Tom, the innovation is now coming from other places. It's the combination of other factors, and that's a big part of what you guys are announcing today, right? Sure, sure, I think one of the things that you see is traditionally it was a fairly homogeneous set of workloads that are getting driven out of the data center, right, and it was, you know, mainstream, DRAM mainstream processors. And you're seeing just an increasingly heterogeneous set of problems that people are trying to solve. And along with that, a more complex and rich set of solutions to go after those problems. You know, just a couple, you know, very common examples if you look at how people are doing machine learning training. Now, very often that is done with a, you know, with a GPU that's getting fed by a graphics memory. If you look at how people are doing machine learning classification, often that's an FPGA, in many cases, fed by, you know, by Hybrid Memory Cube. And so this increasing specialization of workloads and solutions to support those workloads and specialized memories to support that are just an opportunity for, you know, for more value add for a memory company like Micron. So Steve, when we have folks like you guys on, we'd like to talk about, you know, what's changing. And you may gave a stat today about the thousand to one ratio of the efficiency of moving data versus compute. And what are the implications that that has on applications, IT in general, and how do you rearchitect IT to solve that problem? Well, so first, it's hard to rearchitect actually the computing architecture because there's billions of dollars worth of legacy software out there that you still need to run. But one of the big things we can do is, and you know, this is what we're doing at Micron is, if the data is critical, it stays resident and you move computing to the data. So you try to move as much of what you can do in terms of processing that data closer to the data so you're not, you know, sending that all the way over because in a lot of cases, the processors, most of them are increment operations so it'll take a big chunk of data, add one and send it back. And that's just a tremendous waste of time and energy and you can do a lot of that work closer and closer to memory. Eventually, can you get a core in memory? I believe so, it won't run at several gigahertz but you can get that kind of capability in memory and do a lot more of that functionality. And if you can leverage the legacy software, it can take advantage of that and then over time start to build on top of that with new applications, the industry will start to move. And that's, you know, I mentioned the two-summer Olympic cycle. It takes, when you come up with a, one of the reasons why I believe Intel was so successful was like with the 386, it didn't have any 32-bit software and it was a 32-bit architecture but it ran 16-bit code very well. And so as you start to add new features and new capabilities, as long as you can run that legacy code and people see an impact, they'll buy, the hardware will become more pervasive and then the software ecosystem will start to follow. So a question on where new systems are going because clearly this fabric changes an awful lot of stuff. But there's a lot of other things going on as well. There's PCI Gen 4 that's coming along. There's, as you said, GPUs, there's FPGAs. There's a lot of stuff in there that you can make more efficient in that way. And one of the areas that looks very interesting to me is things like NVLink, which can allow you a way of sharing data, of sending data once or sharing that data across multiple different CPUs or the similar CAPI link. So with this host of different pieces in that architecture, how are you seeing the NVME over fabric fitting into all those different pieces? How are people gonna put these together? Do you wanna take that or? I'll let you take a stab at it. Okay. Well, you know, most of the architecture, I mean, NVLink is a proprietary architecture. It's a great architecture, by the way, and with people that know how to do really high-speed links, and Diddy is one of the best at it. But a lot of these architectures like PCI and like some of the others that are coming up, they have a lot of legacy baggage, and so you're still having to continue with either the software paradigm or whatnot. I actually do see, so NVME over fabric is a good step, but I actually see that at some point in time, there will be an abstracted interface that will become the network interface. And whether that becomes some variant of NVLink, OpenCAPI, CCIX, GenZ, that becomes the next abstracted interface, but it has to have networking capabilities because if what we are planning to do actually comes to fruition, you will have the compute memory subsystem there, and the only connection between these components will be through a fabric. And when you want it to be a switchless fabric inside the rack and then outside the rack, you wanna go to the top of rack switch, but you wanna make it as lower latency as possible to the other elements that it's gonna connect to. Tom, I wonder if we could talk about some of the broader trends in the server and networking space, how they're affecting your business. So obviously you got the hyperscale guys, you've got certainly consolidation in the server market in terms of the number of players, Dell buying EMC and other interesting sort of piece of the consolidation. On the networking side, we're pushing bottlenecks to the networks, things like traffic's going east-west versus north-south, still Cisco's a dominant player, but you see others, whether it's software defined or folks like Arista really taking a stab at the leader, what do you see in those businesses and how is it affecting your business? Sure, well let me take two, one is just kind of growth in the scale of the data center opportunity and then I'll talk a little bit about how we see the cloud or hyperscale provider is complimenting the traditional still very strong OEM server partnerships that we have. From a scale perspective, obviously you heard Darren and others earlier today talking about a lot of the factors that are driving data scale growth be that public, private, hybrid cloud from video capture and distribution that's upsetting the TV industry as we once knew it to the support for the myriad of IoT devices that are out there through the opportunity to monetize data whether that's consumer facing public data or industrial data that has never been digitized before. And so all of that is what's behind the fact that you go back three years ago and the PC was the largest consumer of DRAM bits not only within our business, but for the company. Mobile passed that in 2015. Now the data center as a whole is larger than that and if you look out to the end of the decade the cloud and enterprise each individually will be larger than PC. So it's a fairly fundamental shift in terms of the volume that's driving our business. I think back to the question of kind of the complimentary nature of the traditional survey OEMs and the hyperscaler cloud providers. Again, I'd mentioned both of them are seeing that ever increasing complexity and heterogeneity of workloads and opportunities to get in and understand how differentiated solutions can support those better. It's just a great opportunity for for a memory provider to add value. And that's true across both of those classes of players. If there's one thing I think that we're seeing from the hyperscale players they tend to be able to have a more controlled set of workloads and a more controlled operating environment which can tend to make their validation problem a little bit simpler and more straightforward. And so when there are new technologies that come in to the extent that they can validate those a little bit quicker, they can adopt them faster and so can perhaps be a little bit earlier adopter on that curve. Right, Steve I wonder if we could talk a little bit about this influx of data and the implications for systems and memory architecture. I mean it goes back to the mainframe days of MVSXA and expanded memory and we were thinking about okay how do we persist beyond this limited resource didn't have nearly the data volumes that we do today. And then you mentioned Fusion IO in your remarks. We were kind of David Flynn groupies back in the day and the whole idea of atomic rights and eliminating scuzzy. So what are you seeing in terms of architectures changing to accommodate this exponential data growth? Well, so I think there's really a couple of things. When we talk about hey we're gonna have all this data and where we're gonna store it. A petabyte of data or an exabyte of data so Darren talked about zettabytes of data. Exabyte of data in flash is roughly about $172 million if you look at the current cost per gigabit equivalent. So I'm not really sure you can store all that data. What you're gonna have to do is find the data that's relevant. For example if you're taking a picture of an image and that image is the same 24 hours of the day but when a cat moves through you wanna see what changes. That's the kind of information you really wanna change. So I think what's gonna happen is there's a tremendous amount of data being generated but there's gonna have to be a lot of filtering or action at the edge to decide what data to keep and what data not to keep. And I think that's where a lot of the big innovation is gonna occur because you just don't have the backbone. 5G implementations will certainly add a lot in terms of a more robust, reliable cellular capability but it's still not gonna be able to provide the kind of backhaul you need if you're just sending raw data to the data center. There has to be significant amount of intelligence out at the edge and that's where energy efficiency comes into play because not everybody has a power cord they can plug into. So the thinking of how this architecture is gonna evolve is gonna be huge but it all has to revolve around what data do you keep, what data do you throw away? So it's either gonna be disposed of at the edge or maybe persisted at the edge but it's certainly not gonna be shipped all back to the centralized. It's kind of like the JPEG or MPEG. MPEG keeps a frame and then they keep data on terms of how that frame changes so they don't continue to maintain data on every frame. It's just how those frames happen to change so that that's how they can get the compression ratios that they're looking for. I think we're just gonna have to apply that kind of capability across the broader spectrum. Well what that means is you're gonna have to have storage out there enough to keep the data and then be able to have the compute at the edge enough of it to be able to say what's relevant and what's not. And the algorithms have to change constantly. Right. So I mean that's a very interesting area of the edge. The amount of compute that we're gonna be pushing out to the edge is in the opposite direction to currently where everything is going towards the cloud. That seems to be a very strong move and a necessary move to actually be able to then bring the nuggets back up. But you know I don't think we're gonna be pushing 64-bit double precision arithmetic out to the edge. I think we're gonna make it as simplistic as possible to do whatever work actually makes sense. Right. Do you see the devices at the edge though it's starting to put a lot of this processing in for example camera putting it into the camera itself and those architectures being changed quite dramatically to be able to solve some of these data reduction problems as early as you possibly can. Yeah, I mean I think again I think there's one factor which Steve has already covered very well which is simply the raw bandwidth will not be there to bring everything back into a centralized location no matter how many data centers you build out around the world. But I think the other factor which won't impact all applications but a bunch of ones that matter is gonna be latency. Which is you have an advanced driver automation system in a car and you need to be making real time decisions. And so I think the combination of those two is gonna continue to be a driving force towards that intelligence at the edge. At the same time I don't think no matter how phenomenal a job we do at that edge processing filtering compression, call it what you will that we're gonna see a material drop in the demand for both memory and storage sitting centrally. I think it's gonna barely keep up as we push that process into the edge. Totally agree. That wasn't implying that you would get rid of the cloud it was just that on top of it you're gonna have to push in a huge amount of connectivity and I was just is that gonna be a separate type of product or are you gonna take your architecture and adapt it to different spaces? You know I think it's again it's gonna be increasingly as volumes grow it'll be increasingly specialized. I mean I think that so you wanna put intelligent agent functionality into a smartphone. The requirements of that and the ability to perhaps over time have a more hybrid solution where you're perhaps knocking off the easy bits of that problem in the handset and sending the tougher ones back back more centralized works. Again to use the automotive example anything that's real time I'm pretty sure the drivers at least in the USA with our legal system we're gonna want that locally. Do you have a point? I completely agree I think because with the stuff that we would actually put in the devices out at the edge will actually add area and cost that the people in the data center may say I'm not paying for that and if you give it to them for free you end up giving and you don't make any money. So I think they will be more specialized. But you know one of the epiphanies I had coming to Micron was just how much we could actually do in the memory array. You mentioned key value store. Yeah key value store is a perfect example in doing machine learning classification. I can actually see building deep neural nets out of these devices because we will have the capability. Training will occur somewhere else but once we've got those trained weights we can actually run those devices. Pretty well. All right we got to leave it there. Gents thanks for coming on theCUBE. We really appreciate you. Thank you for having us. You're welcome. All right keep it right there everybody. We'll be back to wrap up after this short break. This is theCUBE we're live from Micron Summit in New York City. Right back.