 Hey, welcome back everyone to theCUBE's coverage of AWS re-invent 2021 in-person, it's a live event, physical in-person, also virtual hybrid. So a lot of great action online. Check out the website, all the videos are there on theCUBE as well as what's going on here. But all the action's on site in theCUBE's here. I'm John Furrier, your host with Dave Vellante, my co-host founder. We've got David Brown, VP of Elastic Cloud Compute, Compute Cloud EC2, known as EC2. The bread and butter, our favorite part of Amazon. David, great to have you back on theCUBE in person. Well John, it's great to be back. That's actually the first time I've been on theCUBE in person as well. A lot of virtual events with you guys, but it's amazing to be back at re-invent you. We're so excited for you. I know Matt Garmin and I have talked in the past, we've talked in the past, EC2 is just an amazing product. It's always been the core block of AWS. More and more action happening, and developers are now getting more action in there as well. We wrote a big piece about it. What's going on? The silicon's really paying off. You've got the also general purpose Intel and AMD, and you've got the custom silicon all working together. What's the new update? Give us a scoop. Well John, it's actually 15 years of EC2 this year, and I've been lucky to be on that team for 14 years. And so incredible to see the growth. It's been an amazing journey. The thing that's really driven us, two things. One is supporting new workloads. And so what are the workloads that customers have available out there trying to do on the cloud that we don't support and launch new instance types? And that's the first thing. The second one is price performance. How do we give customers more performance at a continuously decreasing price year over year? And that's just driven innovation across EC2 over the years with things like Graviton, all of our inferential chips are custom silicon, but also instance types with the latest Intel, Ice Lake CPUs, latest Milan, we just announced the AMD Milan instance. It's just constant innovation across the ever increasing list of instances. So it's super exciting. So instances become the new thing. Provision in instance, spin up an instance. Instance becomes, and you can get instances, flavors almost like flavors, right? Yeah, take us through the difference between an instance and then the EC2 itself. That's correct. Yeah, so we actually have, by end of the year, right now we have over 475 different instances available to you, whether it's GPU accelerators, high performance computing instances, memory optimized, just the enormous number. We actually hit 500 by the end of the year. But that is it. I mean, customers are looking for different types of machines and those are the instances. So the custom silicon, it's one of the most interesting developments. We've written about it. AWS's secret weapon is one of them, right? I wonder if you could take us back to sort of the decision points and the journey. You know, the Interpurn acquisition, you started working with them as a partner and then you said, all right, let's just buy the company. And then now you're seeing the acceleration, your time to tape out is way, way compressed. Maybe, what was the catalyst and maybe we can get into where it's going? Yeah, absolutely. Super interesting story because it actually starts all the way back in 2008. In 2008, EC2 had actually been around for just a little under two years. And if you remember back then, everybody was like, well, virtualized and hypervisors, virtualization, would never really get you the same performance as what they were calling bare metal back then, right? As everybody's looking at the cloud. And so we took a look at that. And I mean, networked latencies, in some cases with hypervisors, were as high as 200 or 300 milliseconds. And it was a number of real challenges. And so we knew that we would have to change the way that virtualization works and get into hardware. And so in 2010, 2011, we started to look at how could I offload my network processing, my IO processing to additional hardware. And that's where we delivered our first Nitro card in 2012 and 2013. Where we actually offloaded all of the processing of network to a Nitro card. And that Nitro card actually had an Anapuna ARM chip on it. Our Nitro one chip. For the offload. The offload card. Yeah, and so that's where my team started to code for ARM. We started to work on how Linux works for ARM. We actually had to write our own operating system initially, because there weren't any operating systems available we could use. And so that's where we started this journey. And over the years, when we saw how well it worked for networking, we said, let's do it for storage as well. And then we said, hey, we could actually improve security significantly. And by 2017, we'd actually offloaded 100% of everything we did on that server to our offload cards. Leaving 100% of the server available for customers. And we're still actually the only cloud provider that does that today. Just to interject, in the data center today, probably 30% of the general purpose cores are used for offloads. You're saying 0% in the cloud. On our Nitro instances, so every instance we've launched since 2017, our C5, we use 0% of that central core. And you can actually see that in our instance types. If you look at our largest instance type, you can see that we're giving you 96 cores. And we're giving you, in our largest instance, 24 terabytes of memory. We're not giving you 23.6 terabytes because we need some. It's all given to you as the customer. So much more efficient. Much, much more efficient. Much better price performance as well. But that ultimately, those Nitro chips, we went through Nitro 1, Nitro 2, Nitro 3, Nitro 4. We said, hey, could we build a general purpose the way a server chip could we actually bring on into the cloud? And in 2018, we launched the A1 instance, which was our Graviton 1 instance. And what we didn't tell people at the time is that was actually the same chip we were using on our network card. So essentially it was a network card that we were giving to you as a server. But what it did is it sparked the ecosystem. That's why we put it out there. And I remember before launch, she was saying, is this just going to be a university project? Are we going to see people from big universities using ARM in the cloud? Was it really going to take off? And the response was amazing. The ecosystem just grew. We had customers move to it and immediately begin to see improvements. And we knew that a year later, Graviton 2 was going to come out. And Graviton 2 was just an amazing chip but continues to see incredible adoption. 40% price performance improvement of other X86 instances. This is worth calling out because I think that example, the network card, I mean, innovation can come from anywhere. This is what Jassy always would say, is do the experiments. Think about the impact of what's going on here, right? You're focused on a mission. Let's get that processing at the lowest cost, pick up some workload. So you're constantly tinkering and we're tuning the engine. New discovery comes in. Nitro is born. The chip comes in. But I think the fundamental thing, and I want to get your reaction to this is we put this out there on our post on Sunday. And I said, in every inflection point, I'm old enough, my birthday was yesterday, I'm old enough to know that. I saw that. I'm old enough to know that in the 80s and other client servers shifts, every inflection point where development changed, the methodology, the mindset or platforms changed, all the apps went to the better platform. Who wants to run their application on a slower platform? And so in those inflections, so now that's happening now, I believe. So you got better performance and I imagine that the app developers are coding for it. Take us through how you see that because, okay, you're offering up great performance for workloads. Now it's cloud workloads, that's almost all apps. Yeah, can you comment on that? Well, it has been really interesting to see. I mean, as I said, we were unsure who was going to use it when we initially launched and the adoption has been amazing. Initially, obviously, it's always, you know, a lot of the startups, a lot of the more agile companies that can move a lot faster, typically a little bit smaller. They started experimenting, but the data got out there, right? That 40% price performance was a reality and not only for specific workloads, it was broadly successful across a number of workloads. And so, you know, we actually just had SAP, who obviously is an enormous enterprise, supporting enterprises all over the world, announced that they're going to be moving the S4 HANA cloud to run on Graviton 2. It was just phenomenal. And we've seen enterprises of that scale and game developers, every single vertical, looking to move to Graviton 2 and get that 40% price performance. Now we have to, as analysts, we have to say, okay, how did you get to that 40% and you have to make some assumptions, obviously. And it feels like you still have some dry power when you look at Graviton 2. I think you were running it. I don't know if this, it's speculated anyway. I don't know if you guys, it's your data, two and a half, 2.5 gigahertz. I don't know if we can share what's going on with Graviton 3, but my point is, you had some dry powder. And now you have, with Graviton 3, quite a range of performance, because it really depends on the workload. Maybe you could give some insight as to that. And what can you share about how you tuned Graviton 3? When we look at benchmarking, we don't want to be trying to find that benchmark that's highly tuned but then put out something that is, hey, this is the absolute best we could get it to and that's 40%. So that 40% is actually just on average. So we just went and ran real world workloads. And we saw something with 55%, we saw something with 25, depending on what it was. But on average, it was around the 35, 45%. So at least 40%. And the great thing about that is customers come back and say, hey, we saw 40% in this workload. It wasn't that I had to tune it. And so with Graviton 3 launching this week, available in our C7G instance, we said 25%. And that is just a very standard benchmark in what we're seeing. And as we start to see more customer workloads, I think it's going to be incredible to see what that range looks like. Now Graviton 2 for single threaded applications, it didn't give you that much of a performance boost. And that's what we meant by cloud applications, generally multi-threaded. And Graviton 3, that's no longer the case. So we've had some customers report up to 80% performance improvement of Graviton 2 to Graviton 3 when the application was more of a single threaded application. So we started to see. I mean, try another single thread. The other key point I want to add is the time to market is compressing, right? So you have that, go ahead, sorry. No, no, I always want to add one thing among the difference in single and multi-threaded applications. A lot of legacy are single threaded. So this is kind of an interesting thing. So the mainframe migration stuff, you start to see that. Is that where that comes in the whole? A lot of the legacy apps, but also even some of the newer apps, right? If a single threading in like video transcoding, for example, is all done on a single core. It's very difficult. I mean, almost impossible to do that in a multi-threaded way. A lot of the crypto algorithms as well, encryption and cryptography is often single core. So with Graviton 3, we've seen a significant performance boost for video encoding, cryptographic algorithms, that sort of thing, which really impacts even the most modern applications. So that's an interesting point, because now single threaded is where the vertical use cases come in. It's not like more general purpose OS kind of things. And Graviton's already been very broad. I think we're just knocking down the last few verticals where, maybe it didn't support it and now it absolutely does. And if an ISV then ports, like an SAPs ports to Graviton, then the customer doesn't see any, I mean, they're going to see the performance difference, but they don't have to think about it. They just say, I choose that instance and I'm going to get better price performance. So we've seen that from our ISVs. We've also been doing that with our AWS services. So services like EMR, RDS, ElastiCache, have all been moving and making Graviton 2 available for customers, which means the customer doesn't have to do the migration at all. It's all done for them. They just pick the instance and get the price performance benefit. So, yeah. I think, oh no, that was serverless. Sorry, I haven't done that. Well, Lambda actually just did launch on Graviton 2 and I think they were talking about a 35% price performance improvement. Who was that? Sorry. Lambda. A couple of months ago. So what does an ISV have to do to port to Graviton? Yeah, it's relatively straightforward. And this is actually one of the things that has slowed customers down is the, wow, that must be a big migration, right? And that ecosystem that I spoke about is the important part. And today, with all the Linux operating systems being available for ARM, running on Graviton 2, with all of the container runtimes being available, and they saw the open source applications and ISVs being available, it's actually really, really easy. And we just ran the Graviton 4-day challenge and we did that because we actually had an enterprise migrate one of their largest production applications in just four days. Now, I probably wouldn't recommend that to most enterprises, obviously it was a little too fast, but they could actually do that. It's just from a numbers table. That's insanely amazing. I mean, when you think about four days. And when we talked virtually last year, this year, I can't even remember now, you said, we'll just try it. That's right. And so I presume a lot of people have tried it. Well, that's my advice. It's the unknown, it's the what will it take. So take a single engineer, tell them, give them a time. Say you have one week, get this running on Graviton 2. And I think the results are pretty amazing. Very surprising. You know, we were one of the first, if not the first to say that ARM is going to be dominant in the enterprise. We know it's dominant in the edge. And when you look at the performance curves and the time to tape out, it's just, it's astounding. And I don't know if people appreciate that relative to the traditional Moore's Law curve. I mean, it's astounding. And then when you combine the power of the CPU, the GPU, the NPU, kind of what Apple does in the iPhone, it blows away the historical performance curves. And you're on that curve. I wonder if you could explain that to folks. So with Graviton, we're optimizing just across every single part of AWS. And one of the nice things is we actually own that end to end. So when it starts with the early design of Graviton 2 and Graviton 3, and we're obviously working on other chips right now, we're actually using the cloud to do all of the electronic design automation. So we're able to test with AWS how that Graviton 3 chip's going to work long before we've even started taping it out. And so those workloads are running on high-frequency CPUs on Graviton. Actually, we're using Graviton to build Graviton now in the cloud. The other thing we're doing is we're making sure that the Annapurna team that's building those CPUs is deeply engaged with my team who are going to ultimately go and build those instances so that when that chip arrives from tape out, I'm not waiting nine months or two years, like would normally be the case, but I'm actually have an instance up and running within a week or two on somebody's desk starting to do the integration. And that's something we've optimized significantly to get down. And so it allows us to get that iteration time. It also allows us to be very, very accurate with our tape outs. We're not having to go back with Graviton. They're all A1 chips. We're not having to go back and do multiple runs of these things because we can do so much validation and performance testing in the cloud ahead of time. This is the epiphany of the R model. It really is. It's a standard. When you send it to the fab, they know it's going to work, right? You hit volume and it's just amazing. No fab. Yeah, hell of it. Well, this is a great thread. We'll stay on this because Adam told us when we met with them for re-invent that they're seeing a lot more visibility into use cases at the scale. So the scale gives you an advantage on what instances might work, whatever. It makes the economics work. Makes the economics work, hence the timing, the shrinking time to market. Not there, but also for the apps. Talk about the scale advantage you guys have. Absolutely. I mean, the scale advantage of AWS plays out in a number of ways for our customers, right? The first thing is being able to deliver highly optimized hardware. So, we don't just look at the Graviton 3 CPU. You're speaking about the core count or the frequency, and Peter spoke about that in his keynote yesterday, but we look at how does the Graviton 3 CPU work with the rest of the instance? What is the right balance between the CPU and memory, the CPU and the hard drive? What's the performance in the drive? We just launched the Nitro SSD, which is now we've actually, we're building our own custom SSDs for Nitro, getting better performance, being able to do updates, better security, making it more cloudy, which is saying we've been challenged with the SSD in the past. The other place that scale's really helping is in capacity. Being able to make sure that we can absorb things like the COVID spike, right? Or the stuff we've seen in the financial industry with just enormous demand for compute. We can do that because of our scale. We're able to scale. And the final area is actually in quality. Because I have such an enormous fleet, I'm actually able to drive down AFR, so annual failure rates. And we're well below what the mathematical, theoretical, tenant or possibility is. So if you look at what's put on that actual sticker on the box that says you should be able to get a 4% AFR, at scale and with focus, we're actually able to get that down to significantly below what the mathematical entitlement was actually being. That's incredible. What a great, and this is the advantage, and that's why I believe anyone who's writing applications that includes a database, data transfer, any kind of execution of code will use the stack. Why wouldn't they? Why wouldn't they? If you're, we've seen this, like you said before, whether it was PC, then the fastest Pentium, or somebody who's developing a Unix box, right? ISVs wanted to run as fast and as cheaply as possible. And now power plays into it as well. Yeah. Well, we do have, I mean, I agree with what you're saying. We do have a number of customers that are still looking to run on X86. Obviously customers that run Windows. Windows isn't available for ARM, and so that's a challenge. They'll continue to do that. And you know, the way we do look at it is, Moore's Law kind of died out on us in 2002, 2003. And what I'm hoping is, not necessarily bringing Moore's Law back, but that we say, you know, let's not accept the 10% to 15% improvement year over year. There's absolutely more we can all be doing. And so I'm excited to see, you know, where the X86 world's going. I know they're doing a lot of great stuff. Intel Ice Lake's looking amazing. Milan is really great to have on AWS as well. Well, I think it's a fair point, because you look what Pat's doing at Intel, and he's remaking the company. I mean, I've said he's going to follow in the ARM playbook in my mind a little bit, and which is the right thing to do. So competition is a good thing. Absolutely, yeah. We're excited for you and it's great to see Graviton and you guys have this kind of inflection point. We've been, you know, trying it for a while, but now the world's starting to see it. So congratulations to you and the team. Thank you. Just a couple of other things. You guys have some news on instances. Talk about the deprecation issue and how you guys are keeping instances alive real quick. Give the plug. Yeah, you know, we're super customer obsessed at Amazon and so it already drives us. And one of the worst things for us to do is to have to tell a customer that we no longer support in a service. You know, we recently actually just deprecated the EC2 classic network. I'm not sure if you saw that. And that's actually after 10 years of continuing to support it. And the only reason we did it is we have a tiny percentage of customers still using that from back in, you know, 2012. But one of the challenges is obviously instance hardware eventually will ultimately time out and fail and have hardware issues as it gets older and older. And so we didn't want to be in a place in EC2 where we would have to constantly go to customers and say, that M1 small, that C3, whatever you were running, it's no longer supported, please move. That's just a tax that customers shouldn't have to do. And if they're still getting value out of an older instance, let them keep using it. So we're actually just announced and reinvented my keynote on Tuesday, the longevity support for EC2 instances, which means we'll never come back to you again and ask you to please get off an instance because we can actually emulate all those instances on our nitro system. And so all of these instances are starting to migrate to nitro. You're getting all the benefits of nitro for now some of our older Zen instances. But also you don't have to worry about that work. It's just not something you need to do to get off an older instance. That's great, that's a great service. Hey, stay on as long as you want. Yeah. When you're ready to move, move. Okay, final question before I know we've got time to, I want to get this in. The global network you guys are known for, AWS Cloud WAN service. Yeah. Give us an update on what's going on with that. So Vernick just announced that in his keynote. And you know, over the last two to three years or so, we've seen a lot of customers starting to use the AWS backbone, which is extensive. I mean, you've seen those slides in Virtus Keto. It really does span the world. I think it's probably one of the largest networks out there. Customers starting to use that for actually their branch office communication. So instead of going provisioning their own international MPLS networks and that sort of thing, they say, let me onboard to AWS with VPN or Direct Connect and I can actually run the AWS backbone around the world. Now doing that actually has some complexity. You've got to think about transit gateways. You've got to think about those interregion peering and AWS Cloud WAN takes all of that complexity away. You essentially create a Cloud WAN, connecting to it with VPN or Direct Connect. And you can even go and actually set up network segments. So essentially VLANs for different parts of the organization. So super excited to get out there. I think we've seen more customers. The ease of use is the key there. Massively easy to use. And we have, you know, 26 SD-WAN partners. We've even partnered with folks like Verizon and Swisscom in Switzerland for Telco to actually allow them to use it for their customers as well. And the other ones. You can use the service someday when we have a global rollout. Yeah, let's do that, yeah, the Q-global. And then the other ones, the M1 EC2 instance, which got a lot of applause. Absolutely, absolutely. Yeah, I think it's based on A15, which is, you talked about 15. Yeah, so that's the Mac. We've got to be careful, because M1 is our first instance as well. Yeah, right, that's a little decision there, right? So it's the Mac. The EC2 Mac instance with the M1 silicon from Apple, which we see pretty excited to put out there. Awesome. David Brown, great to see you in person. Congratulations to you and the team and all the work you guys have done over the years. And now that people starting to realize the Cloud Platform, the compute just gets better and better. It's a key part of the system. Thanks for sharing. Thanks John, it's great to be here. Thanks for sharing. The silicon angle is here. We're talking about custom silicon here on AWS. I'm John Furrier, David Brown. You're watching theCUBE, the global leader in tech coverage. We'll be right back with more coverage from re-invent after this break.