 And we're here at Computex 2016 and yesterday you had the press conference for Thunder X2 So who are you? I'm Gopal Hegde, I'm the VP GM for the data center processor group at Cavium And this is Thunder X2, this is an upgrade for the biggest, most powerful ARM processor in the world, right? Yes So how much faster is it now? So Thunder X2 is roughly 2-3x higher performance compared to Thunder X1 How can we do that? We can do that by, essentially what we are doing is, we are redesigning the core, processor core used in Thunder X2 from scratch Where we are adding a number of new capabilities such as fully out of order execution, it's a multi-issue core We are using the 14 nanometer FinFET technology which allows us to improve overall frequency So we are targeting a frequency of 2.4 to 2.8 GHz in normal mode, 3 GHz in turbo mode of operation So a combination of a higher core count, Thunder X2 is 54 cores compared to Thunder X which is a 48 core Higher frequency and fully out of order multi-issue next generation core gives us the ability to deliver roughly 2-3x the performance of Thunder X For a wide variety of applications with Thunder X2 So now it's 54 cores per SOC? Yes You can still have 2 or more maybe on the board? You can have 2 SOCs per board interconnected using our next generation processor interconnect, KVM coherent processor interconnect technology That delivers 2.5x higher coherent bandwidth between the 2 sockets compared to Thunder X So there's a faster interconnect technology Yes Where is it located? It's the interface between the 2 processors of Thunder X which allows to create a more of a cache quadrant architecture in a dual socket mode So it's a chip that's somewhere else in the PCB is not an SOC? No, it's integrated on Thunder X, you don't need any other chip Not only that, Thunder X also integrates all the IO required for data center whether it is network IO We have next generation of network IO 25 gigabit so we support 25, 50, 100 gigabit And in addition to supporting 10 and 40 gigabit which Thunder X1 supported We have SATA interfaces, we have PCIe so everything you need to build a next generation server for the data center Thunder X2 integrates in a single piece of silicon So it's custom ARM design right? It's a custom core designed by KVM yes So what did you learn from Thunder X to Thunder X2 in the way you make custom ARM? So what we learned with Thunder X1 has been very successful in a set of scale out application Where application set can be highly parallelized so that can take advantage of our 48 cores That can take advantage of our integrated network and storage IO And applications that can benefit from our integrated accelerators We get significantly better performance What we learned is that there were a set of applications that require high single thread performance And we knew going into Thunder X1 that was not our target set of applications With Thunder X2 we widen our application base to applications that can benefit from high single thread performance Of Thunder X1 applications like web search, real time analytics, data warehousing type of applications Massively parallel databases and things like that So the parallel stuff is quite important and you have so many cores right? Yes And you add more SOCs on the board and you have a rack with hundreds of SOCs And you have to write server software that's very good with parallel stuff No But now it's not as much required? No, no, so basically the way it works is that multi core architecture has been around in the industry for a long time So you have Intel CPUs today in the market that support 12 cores, 14 cores, 18 cores, 20 cores, 24 cores, etc So the software infrastructure can take advantage of the multiple cores that are in a processor So we don't have to do anything new from a software development standpoint to take advantage of multiple cores And all that is leveraged on Thunder What Thunder does is that we take the core count to another level We have a lot more cores that allows applications to scale even more than what is possible with today's CPUs out there in the market So this one, the Thunder X2 now is even more ready for the cloud? Thunder X2 improves performance for all the applications, Thunder X1 services today Whether it is cloud storage, analytics, cloud compute type of applications, etc In addition to all those applications, we now broaden the set of applications that we can service with Thunder X2 The applications that require high single thread performance which were not serviceable by Thunder X1 So all those companies actually have a highly scale out workloads that are parallelizable with Thunder X1 We can actually service a lot of those workloads and we are into a number of customers who are actually taking advantage of the capabilities in our current Thunder X processor for a wide variety of workloads However, these customers also have workloads like for example, one of the large workloads at Google is the web search application That we can actually target much better with Thunder X2 However, there are workloads like storage and analytics, etc Which we can service today, which we are servicing today with Thunder X1 Do you feel the earthquake? Just a second We're going to resume It's moving a little bit It's normal, it's every day here, right? Oh wow Okay, cool, it's moving So, what's it called? You were saying that... How is the adoption of Thunder X1 Has it been like a test chip for Thunder X2? No People are testing? Are they really deploying massively? Yes, so if you look at Thunder X1, Thunder X1 has a lot of market traction today in the market We have well over 20 different ODM platforms If you go inside, you look at platforms We have 1U, 2U, Rackmount, Blade, density optimized platforms based on Thunder X1 Which are developed by leading ODMs and OEMs here So if you look here, you have platforms on this wall Can we check this one out? What is this? So this is a platform for high performance computing Developed by a company called Eco Computer Engineering in Europe What you see here is a Tesla GPU which is integrated here You have an infini band adapter This is for high performance computing Application based on Thunder X It's a dual socket Thunder X platform What's the advantage of combining the NVIDIA GPU with your arm? So NVIDIA GPU delivers very good floating point performance And it's very important for high performance computing applications So for high end HPC applications combining Thunder X capabilities Which is more of a scale out capabilities with a GPU Which gives you ability to deliver a very high high performance computing type performance For floating point applications, you get the best of both worlds So with Thunder X2 and this, it's going to be 2.5 times faster? Yes, absolutely Everything here is going to be 2.5 or 3 times faster, right? Yes, absolutely Let's check some more examples So this one here, you have multiple platforms on Gigabyte This one here is a Gigabyte platform It's a dual socket platform What you call in the EATX form factor This is a very popular form factor One of the ones, this form factor, a dual socket Is roughly 12 hard drives here Or in a 3.5 inch hard drive configuration Or 24 hard drives in a 2.5 inch configuration Is very popular for a lot of Hadoop type of deployments Like big data type of deployments Cloud storage deployment, cloud compute type of environment It's a very popular form factor This one here, this is a different version This one has 12 3.5 inch hard drives This one has 24 2.5 inch hard drives One of the benefits of Thunder X, what you will notice is that The customer has to just redesign the motherboard The chassis, the power supply Everything is what they use today in the market And that enables them to get to market quickly With lower R&D spend And also be able to leverage the cost Of a high volume chassis power supply infrastructure Which makes the overall platform very cost effective For our end customers How hard is it to design those motherboards? It's actually very straight forward Typically our customers go from design to production In about 6 months So it includes the complete design The bringing up the board Running through full testing And then running the production IBM provides full set of collateral Design guides, tools, etc In order for the ODMs to go through this process In a very efficient manner As you can see a lot of these guys have done Taken several platforms based on Thunder to production This one here is a platform With the GPU from Gigabyte What you see here is This is the G2-2060 Which is the GPU server from Gigabyte What you see here is a set node from Gigabyte What you notice is just one motherboard And you have a 4, 10 gig, 140 gig interface There is no NIC card here All the ports are built on to the chip There is no HBA here All the storage interfaces are built on Thunder X So you don't need any separate HBA And this is a very cost effective node For cloud storage applications It's a very popular workload on Thunder This one is essentially a one you one socket Thunder platform Very low cost, it's pretty popular in the cloud Pretty popular, what do you mean by pretty popular? So people deploy this to Just to provide cloud virtual machine access 48 cores, you can support a large number of VMs Integrated networking, integrated storage What it means is that each VM can have its own NIC And each VM can have its own hard drive If that is what is needed So these are platforms from Inventec Inventec is one of the leading podiums Here, based on Thunder And these are two different contractors So is Thunder X2 going to be pin to pin compatible? No, it's not going to be pin compatible Because it's based on the 14 nanometer FinFET technology However, we are taking all the learnings In terms of design collaterals, design guides, tools Everything we have developed for Thunder X1 We will be leverageable for Thunder X2 as well Which simplifies how long it takes for customers To go to production with a Thunder X2 platform What is this one? This one is what we call a cloud scale rack One of the interesting things that is happening With a lot of our customers is Cavium provides a lot of technology elements for the data center So we have silicon that can be used for top of rack switches What we call Xpliant And it's a full software defined networking switch SVN switch We enable a number of servers based on Thunder X technology Compute servers, storage servers, etc We enable appliances which can be used For security applications, etc So this one is a rack that is fully powered By Cavium technology based on platforms That are developed by our OEM and ODM customers So we call it a cloud scale rack All these platforms actually work together With each other in order to deliver A pretty compelling value proposition to our customers For example, the fabric in our servers Works with the top of rack fabric In order to provide fully optimized Traffic management, quality of service VM level monitoring capabilities All these servers can be provisioned using OpenStack You can run more compute nodes or more storage nodes Depending on what workload is required Our liquid security appliance Provides VM level security and key management It's fully cloud ready And our liquid IO adapter which is in this server Can be used to enable NVMe fabrics And several other offload technologies That are required for the data center So this is right here You call it this cloud scale rack And as you can see it includes All Cavium powered Technology for data center So what you can do is Customers can now Build platforms based on Cavium silicon solutions And deploy entire data center Using those platforms Built by our OEM and ODM partners For their data center application So you provide the whole rack We don't provide the whole rack We provide the silicon We provide the technology All of it All of it, yes So the whole range of server stuff Server, networking Network IO Security Switching, Ethernet switching We provide all those technology building blocks And so is this the best rack in the world And how much better is it going to be with ThunderX2 So with ThunderX2 We deliver Higher performance on the server side We enable a wider range of applications With the ThunderX2 technology So all those will result in And we support next generation IO interconnects Which improve the overall throughput IO throughput for each of the servers So combination of all these Delivers substantial benefits To our customers Cavium is one of the very few companies Which has all the building blocks Required to build an entire data center Which can be powered through Cavium technology So are you jumping from 10 gigabit connections to 100? So we're going from We have multiple interfaces That support 10 gigabit, 25 gigabit 40 gigabit, 50 gigabit All the way up to 100 gigabit IO We support multiples of those interfaces ThunderX2 Yes, today with ThunderX1 We support 10 gig and 40 gig IO All the way up to 100 gig Okay, so we have the ability To take the IO bandwidth IO throughput per server Two and a half times What is available in ThunderX1 With the ThunderX2 products So lots of companies Have publicly or secretly Been testing ThunderX, right? Yes And what have they been asking? Are you going to do everything in ThunderX2 That they've been asking? So a lot of customers have Done a lot of work on Testing a wide variety of applications On ThunderX1 A number of them are actually Deploying the ThunderX1 Based server solutions You see all these platforms Developed by leading OEMs and ODMs They're developing these With a specific one or many End users in mind, right? They're not going to spend their money And develop these platforms Without having the end customer So that is really what is going on here And as they deploy ThunderX based products The way it happens is That they test us for one workload Then they build a rack They test us for more workloads And where we are applicable They start deploying that solution There were some applications Many applications Which were more of a scale out data center type Where ThunderX1 is highly suited for And what we are trying to do Is with ThunderX2 Address a broader set of applications Especially applications that require High single thread performance And more memory capacity And higher memory speeds So high single thread performance Is a big deal High single thread performance Is a big deal for some applications It's not required for all applications And as we have And that is what we are trying to Target with ThunderX2 So how is the power consumption The heat of ThunderX2 Compared to ThunderX1? So the power is going to be In the similar power envelope And we achieve that By going to a 14 nanometer FinFRED technology Which gives us a significant improvement In power for ThunderX2 And in addition to that We have integrated a sophisticated Power management technology On to ThunderX2 Which actually allows us to monitor And manage power Based on what the workload requires Optimizing the power for applications That sounds good That sounds like a potential way To even save more power Absolutely How much more power could be saved It depends on the application That's the biggest deal right? Yes it is It depends on the application Depends on what is the peak Power consumed by the application Etc So this looks super nice Super awesome How soon is ThunderX ready for this? I mean when are you launching? ThunderX is already in production And ThunderX has been in production Since Q4 of 2015 And ThunderX based platforms Are shipping in volume production Today you see a lot of them here From leading OEMs and ODMs ThunderX2 will start sampling In Q1 of 2017 So sampling Q1 2017 That's 14nm That means there are some extra challenges And making sure everything is super stable And everything Or you just know exactly how to make it happen Yeah we know exactly 14nm technology is not very new What we have done is We have done several test chips Etc So in order to make sure That we understand the nuances With the technology And 14nm technology Is actually used by other people today To produce products in the market So we are not the very first one Out there with 14nm By any stretch of imagination Is it possible with ThunderX2 Now that Google, Facebook and Amazon Everybody is going to just announce That they are going to deploy Data centers with your chips That depends on them Okay so we sure hope so We expect we have a number of customers In large data centers Who are actually looking at ThunderX1 And we expect the momentum Will continue with ThunderX2 And as you do this awesome stuff The best ARM server chips In the world right What has been Intel's response In the last 2-3 years And are they able to Are you just going to totally overtake them Are you overtaking them totally? No no no I think Intel is a very very large company They have a huge installed base They deliver servers all the way From home all the way To supercomputers right What we are playing Is in a small part of the server market It's scale out compute It's cloud Okay Intel is very strong there We are not head to head Comparing head to head competing with Intel Because we are a very small company We are very focused We are focused on setup applications That we are good at And we are focused on winning Customer by customer workload By workload So ARM is quite bullish About the server market share For 2020 I think they said 20% Or 25% Yes Are you going to be Most of that What is your target How many percent of them servers So we really want to be successful In this space And we hope that ARM is going to get market share As you mentioned in the server space And we hope to be One of the leading vendors Delivering ARM server solutions To the market