 Live from San Francisco, it's theCUBE, covering Micron Insight 2018, brought to you by Micron. Welcome back to San Francisco, everybody. You're watching theCUBE, the leader in live tech coverage. I'm Dave Vellante, he's David Floyer, and we're covering Micron Insight 18. It's all about bringing together artificial intelligence and the memory and storage requirements. We're here on the Embarcadero. We got Treasure Island that way. We got the financial district over there. We got Golden Gate Bridge behind us. Tom Ebi is here as the Senior Vice President and GM of Micron's booming compute and networking business unit. Good to see you, Tom. Great to be here. And Primalav Savla is here. He's the Director of Deep Learning at NVIDIA. Welcome. Thank you. So obviously, some of these new emerging workloads require collaboration between folks like Micron and folks like NVIDIA, but Tom, why don't you kick it off? What are some of the big trends that you're seeing in some of these alternative workloads that's driving this collaboration? Yeah. Well, a lot of what we're talking about here today is the drive of AI and machine learning workloads and the implications for memory. Certainly, there's a host of them. Natural language processing, photo and image recognition, applications in medical research, applications in optimizing manufacturing like we're doing in our fabs, and there's many, many more. And of course, what's exciting for us is that to support those in an optimized way really does require the mating of the optimal processing architecture, things like GPUs, with the right high bandwidth low latency memory and storage solutions. And that's what leads to great partnerships between companies like Micron and NVIDIA. David was explaining at our open the intensity of the workloads that you guys are serving and how much more resources it requires to actually deliver the type of performance. Maybe you could talk about some of the things that you're seeing in terms of these emerging workloads. Yeah, so at NVIDIA we build systems for accelerated computing. AI and deep learning is a very quickly expanding field at this point, which needs a lot of CPU horsepower, right? What we are seeing is that different applications, like you said, whether it's image processing, whether it's video, whether it's natural language processing, the amount of data that is there that is required to do deep learning and AI around it, we break it up into two workflows. One is the training, where you actually train the software and make it intelligent enough to then go and do inference later on so that you can do, get the results out of it at the end of it, right? And we concentrate in this entire workflow, and that's where when we are looking at it from a training perspective, the GPU gives you the processing power, but at the same time, all the other components around it perform at their peak. That's where the memory comes in, that's where the storage comes in, and we need to process that data very quickly. Yeah, so we know from systems design that you got to have a balanced system or else you're going to just push the bottlenecks around, right? We've learned that over the years, but so it's more than just slapping on a bunch of storage and a bunch of memory, right? You're doing some other deeper integration. Is that correct, and what is that integration? Yeah, I think the two companies have had a great relationship. Just to talk about a couple examples, we essentially codefined a technology called GDDR5X, which greatly enhanced the speed of graphics technology. We jointly introduced that to the marketplace with Nvidia about 18 months ago, and then worked with them again very closely on a technology called GDDR6, which is the next generation of even faster technology, and we were their launch and ramp partner for their recently announced G-Force RTX line of cards. And it's a very deeply engaged, early in the process, define the standards, jointly develop the solution, very intimate sharing in the supply chain area. And it's a great relationship for us, and we're excited about how we can continue to expand and extend that relationship going forward. So obviously there's the two parts of that you said, the learning part of it, and the inference part of the computing. What are you seeing is the difference between the two? Obviously at the end of the day, the inference part is critical. That's got to be the very fastest response time. You have to have that in real time. Can you talk a little bit about what you're doing to really speed that up, to make that microseconds as opposed to milliseconds? So from an NVIDIA perspective, we build the entire end-to-end tool sets for training and inferencing. We have a set of libraries that we have made it openly available for all our customers, all our partners and users, so that they can go download it and do the training, so they can use the different frameworks and libraries to accelerate the work that they're doing, and then transform it onto the inference part. We have something called TensorRT, which is basically TensorRealTime that gives the capability to get these answers very quickly. So on our T4 or the Turing chipset that we just announced, we can get a very high performance for our image, so any kind of image recognition or image processing that we need to do, we can do that on those systems very quickly. And we can meet, we build entire architectures, so it's not just about one piece, it's about the whole end-to-end architecture of the system. So we heard earlier today in the analyst briefing, in the press briefing, that Micron certainly in this last 40 years has changed. We're seeing a lot more diversity. Usually it's all about PCs. Now there's just so many alternative workloads emerging. Clearly NVIDIA is playing there as well with alternative processing capabilities. What do you guys see as some of the more exciting, emerging workloads that are going to require continued collaboration and innovation? I think to build a little bit on some of the earlier comments about the need for real-time inference, one of the things in the area of diversity that we found interesting, the relationship between Micron and NVIDIA in high-performance memory really started around their graphics business. But we are seeing in other markets, closer to the edge, in automotive, in networking and in other areas where there's a need for that real-time performance. Yet there's also a need for a degree of cost-effectiveness, perhaps a little more so than in the data center. We're seeing technologies like GDR6 being applied to a much broader range of applications like automotive, like networking, like Edge AI to provide the performance to get that real-time response, but in a form factor and at a cost point that's affordable for the application. Anything you'd add to that, Kamal? So I would also add is you talked about applications, different applications that are changing. Today we announced our new set of libraries and tools for the analytics space. That's again a big workload in the enterprise data centers that we are trying to optimize and accelerate with machine learning. So we announced a whole set of tools which take in these large data sets that are coming in and applying it in this environment, in the data centers, and using it to get answers very quickly. So that's what NVIDIA is also doing is expanding on these capabilities as we're going. And as these components and as these technologies get better, it just gets our answers much more quickly. As execs in this space, and you guys both, you know, you're component manufacturers and so you sell to people who sell to end consumers, how do you get your information in that sort of pull-through? You could have worked, obviously you work with your customers very closely. How do you get visibility onto their customers? You're just going to go to shows, you go do joint sales calls, how does that all work? Yeah, I mean, certainly some of that is in discussions with our customers and their marketing groups about what they're seeing from an end customer point of view. But certainly there's other paths. I mean, one of the reasons behind the $100 million venture fund that we announced today is one of the best ways to get that advanced insight is to be working with some of the most innovative startups that understand what some of those end-user needs might be and are developing some unique technologies. And so there's a range, working with our customers through venture fund and others. But yeah, it's important that we understand those needs because the lead time to developing the solutions, both memory and processing architectures is quite long. Of course, everybody wants to work within the video. You guys are getting inundated. We're the most, we're tighter now. Of course, there's not a lot of choices here when you're talking about the levels of components that you're selling. What's life like at NVIDIA? I mean, they've been knocking down your doors to do partnership. I think we've grown from being just a component of now being a complete system and an architecture. We don't only just build just a chip that the GPU was. We also build full SOCs. We also build the libraries and the software and the tools that are required to make this complete end-to-end solutions. And we also do a lot of open source technologies because we want our customers and our end cost partners to build and take what we have and go beyond what it's capable of. And that's where we add value at the end of the day. Yes, it's all of us together. We need to work together to make that much, much faster as we go forward. Yeah, the tooling is incredibly important. This is complicated stuff. It doesn't just work out of the box, right? And so you need an ecosystem as well. Yes. And that's what you guys have been out building. Tom, we'll give you final thoughts. Yeah, well, I guess to build a little bit. Certainly, NVIDIA is moving up the stack in terms of the ecosystem, the software, the complete solution. And I think Micron does as well. Like you commented, traditionally it was a component play. And increasingly, we're going to be building subsystems in memory and storage that occurs today on the storage side. And I think we'll increasingly see that in memory and with some of the future very promising technologies like 3D crosspoint. Yeah, the gone of the days where you just get piece parts and put them all together. They need you guys to do more integration, more out of the box, like you say, subsystems. So guys, thanks very much for coming to theCUBE. Really appreciate it. Thank you. Thank you. All right, you're welcome. All right, keep it right there, everybody. We'll be back in San Francisco. You're watching theCUBE from Micron Insight 2018. Accelerate Intelligence. We'll be right back right after this short break.