 So here just after the ARM press conference here at Computex 2017 you launched a new GPU Yes, we did the Mali G72. It's our latest high-performance GPU for you know premium mobile So this one has more performance and a whole bunch of new features also and you're talking about machine learning What are you talking about? So this one's primarily about improving efficiency, so we've got more graphics energy efficiency we've got improved execution efficiency for machine learning algorithms and also It's smaller. So we've improved the area efficiency for our silicon partners who spend a lot of silicon putting down our GPU So they care very much about the size. So basically everything's better But with all those newer smaller Nanometers actually isn't there lots of space on the silicon that they need to use and they can just use it up for GPU? No So They used to be about 10 cents per square mil was this it was the going rate for silicon and now it's actually increased from that on the Newer FinFET processes so they care very much particularly if you're putting down a premium mobile GPU You know 15 square mil or something like that That's a lot of silicon and that's a lot of dollars But when you do computer vision the stuff from apical and all that stuff Mm-hmm Is it running on the GPU or is it some other parts of the SOC that do that stuff? So different of our arm partners will do things differently One of the things we've seen is that running machine learning is something that people want to do in different places so the Cortex a 75 that was announced today is also made more efficient at running those algorithms and The Cortex the Mali g72 our GPU has also been improved for executing in those algorithms and the dynamic connection technology which enables people to connect Accelerators coherently is also something that we're you know, we're announcing and it people are using Because they will want to connect their accelerators Not necessarily ones that we've produced but their own in that way So we're just trying to make it easy for the arm partnership to do these process this these type of new algorithms In the way that suits them best for their particular consumer devices So you already did you started beat frost with the g71. So what's what's what's been the The the potential with that what what are you achieving with this? So g71 has appeared in a couple of consumer devices a very high-end Samsung Galaxy S phone and also in the Huawei Mate 9 and we've had great success with that Mali g72 were not announcing licensees for today, but you know the usual suspects and We would expect that you to see that in consumer devices either later this year or very early next year And the bit frosted a whole bunch of Reoptimization of how you do a GP on a on a mobile as you see yeah So Mali g72 as I say it's primarily all about efficiency compared to Mali g71 We've gone to a great deal of optimization trouble from Mali g71 It's still on the by frost architecture, but we've made a whole bunch of optimizations Perhaps 500 little ones and you know three three major ones One around the execution of machine learning algorithms where we've reworked the data data path in the ALU And also made some changes around the instruction cache Configuration and that's given us about 17 percent speed up on machine learning And also we're getting very healthy improvements in energy efficiency graphics Of course most of these premium mobile devices The limiting factor is the thermal budget, so it's very much what performance can you give me within this thermal factor? Different manufacturers will of course have a different number for that Including case design and you know chip packaging and things like that But they will have an absolute number and they will want to Maximize the performance they can get out of that thermal budget So it's possible to crank up the megahertz and stuff like that that helps Yes, so different of arm different of our partners will do things differently So some people will put down lots of cores And then crank down the megahertz run it very slowly and run it at very low voltage So you get the performance by having lots of cores running slowly Some of our partners will put fewer cores down but crank up The megahertz and crank up the voltage and get the performance that way And what we actually see is you get a spectrum of these across the arm partnership people doing things differently Is it true that on the GPU there's more An industry of doing parallel computing that's not maybe as developed on the CPU side So actually you know a whole bunch of cores and it just works the software Yes, I mean in graphics itself is the ideal parallel workload So the way the graphics problems are specified in API terms is things like For every pixel on the screen execute this piece of code for every vertex in the geometry Execute this piece of code and that could be you know eight million pixels. It could be a million triangles So you this huge amounts of inherent parallelism within the code We know we work within a thread pool of thousands and millions of threads Whereas typically in CPU land, you know finding finding numbers of threads to execute can be more of a problem and then of course some computational problems are Thread parallel rich and so executing those on a GPU through one of the compute frameworks the compute API's Tends to be very successful then but running code without lots and lots and lots of threads is Is not good on GPU. That's better on the CPU You're designing by far the most popular GPU in the world, right? It's shipping a billion per year now So in 2016, we are very proud. We were the world's number one shipping GPU Our partners because of course, we don't ship any chips at all But our partners shipped a billion chips containing our GPU. I mean, that's a pretty amazing number You know, I have to pinch myself when I hear that Because it's not that many years. It's been going on, right? Well, it seems like a lifetime to me. So 2005 my boss walks into my office and he says are you busy and I was very keen to impress him in those days So I said no, not too bad. Why what do you want? And he said I want you to go out and buy a graphics company And it actually took me a year and in 2006 we bought phalanx in Norway 25 engineers with a lot of very smart talent and some attitude and We grew that team and the media processing group in arm is now well over 500 people and it's taken us from 2006 to 2017 and Yeah, in 2010 we'd hardly shipped any volume at all 2011 we started shipping in volume and in 2016 we shipped about partners shipped a billion GPUs It's been quite a ride. It really has so there's some of these engineers in Norway, right? Are they just like enjoying life and doing amazing GPU work and how does it work? So off the four founders of the company, we've still got two of them working for us And they're still having a good time in designing GPUs and several of their friends from those 25 Are still with us. I probably a large proportion of the 25 still work for us And yeah, they get up in the morning and they live dream GPU design and we're very very lucky to have them working for us very smart people now as I say You can't build a world-class GPU out of 25 people and you know, we've we've had to supplement them with another 400 So but that's that's the way it goes. We now have people working on the design in San Jose in California in Cambridge in the UK in Lund in Sweden and in Trondheim in Norway and Shanghai in China and So there's self-driving cars there's you name it everybody's using GPUs for all sorts of things You need to accelerate stuff super fast Recognize things absolutely and it's offline recognition and online and everything. Yeah, this is a very important thing machine learning is something that some people are doing on device and some people are doing in the cloud and Certainly we at arm feel that an awful lot of that Machine learning the inference needs to be done on device You need it done locally because you need it done quickly. You don't want latency and You might care about that data. So there's privacy issues about transmitting it over the internet security issues so doing it locally on device At low power with quick latency is what most people are going to want And if if those arm power laptops like my Samsung Crumble Plus if they feel super nice and smooth You have a big role to play in making it nice. Oh, absolutely on the the way in which compute devices is Changing is one of the fascinating things about this job. So, you know, you used to have a very clear classification between You know a phone and a laptop and then along come tablets and then you have phablets and then you have Chromebooks And then you have two in ones and you know, this is this huge variety of devices on which people are doing their Primary compute and I just think that's really interesting people get the devices They want as opposed to the devices that other people tell them they ought to have and There's this gaming also and Android is the platform of the future for gaming So mobile high-fidelity mobile gaming is is going gangbusters. It's growing about twice as fast particularly in China It's growing twice as fast as other forms of gaming. So, you know, the casual gaming What you have on phones is rapidly being supplanted by Really high-fidelity Gaming and the reason for that is the capabilities that we now have on these mobile platforms both in CPU terms But also in GPU terms you have The capability for these these new rendering techniques, you know, multiple render targets things like that Deferred rendering in G buffers, you know, there's a lot of techniques that have come Down from bigger devices onto mobile gaming to improve the quality of the images and To have a more compelling immersive experience on those gaming So there's more and more triple-a games that are coming out on Android the amount of money in this sector is increasing all the time because You know people are writing these games in these in these segments for profit, you know This is a business and there's a lot of money being made in mobile gaming now and all these tools They have an SDKs when they develop high-end games on Xbox latest Xbox PlayStation and all that is coming over to the arm World, right? Yeah, the ecosystem play is very important here And we've spent a lot of time in with in in our Mali working with the game studios and the games engine developers So a lot of gaming is done through this intermediate gaming engine And we've spent a lot of effort ensuring that that's been suitably developed and optimized for use on our Mali some other companies are using other GPUs to do a cloud grid Services does it make sense to use the the Mali GPU to do cloud stuff? There's no reason why not. It's not a primary focus for us at this time We are primarily focused on the very high-volume markets and actually there's a relatively small number of servers in the world whereas there are massive numbers of Portable devices so that's where we're focusing at the moment, but you know We'll move on to those other segments in time What was the announcement today a big deal about being the the VR and AR Chipset now so there's really three things that we're concentrating on in terms of the Mali g72 and They're all about making them better So we've made machine learning processing run faster and more efficiently on our Mali g72 We're looking at mobile VR And the high-quality Rendering effects that we've introduced there some extensions like mobile multi-view We've done there and and also the high-fidelity gaming so you know mobile VR machine learning and High-fidelity mobile gaming these are really the three areas. We're concentrating on because Google announced that they want to have crazy High resolutions in VR, but there's no way to have the bandwidth so you need to optimize Just where you're focusing absolutely. So think of foveated rendering so Right take your thumb and forefinger hold it out at arms length That's the if you look in the middle of that hole That's the bit that you can see in any detail That's the bit you focus on and your brain actually fills the rest of the picture in by you know Strange magic brain stuff. So actually what we do in Foveated rendering is if you can detect where the person is focusing their eye on That's the bit you render in detail and you render the rest of the stuff around it in much lower detail and The brain is just it says yeah, that's fine. It it's all good It says all the the whole quality is good even though actually the bit It's not really looking at is in much lower detail and that's the thing called foveated rendering and we've we've produced some extensions for that in on our GPU and Made that much more efficient and that's that's going really well because that is an optimization that Pays huge huge dividends because the bit that you're actually rendering in full detail is so small compared to the whole of the detail of The screen so you can save an awful lot of power and performance You need to be accurate on detecting where people are looking and You don't need you don't want to see a latency you you want to you want to have a license incredibly important but the tolerance of people from moving their head or moving their eye to the picture in their VR headset changing is varies from really really fast to unbelievably fast, so Certainly you're looking at frame rates, you know up around a hundred frames per second to keep most people happy and Some people are incredibly sensitive to that latency delay and you have to have it even faster Otherwise, they feel sick frankly motion sickness The one of the things I'm most excited about The Google IO was the VPS or what do I think they call it VPS indoor positioning? Yeah, so you're totally ready for this right this basically is project Tango Yeah, is Google max indoor Google maps indoor That's what it should be right should be Google maps when you go indoor Yep, and that's gonna be accelerated absolutely. So if you think about your eye If you think about GPS, you know, it's accurate to within I don't know a few meters But actually if you move your head even half an inch, you know Even a quarter and it's just a couple of millimeters Side to side you can see a huge change in the picture that your eyes are seeing and with stereoscopic vision You can actually work out where you're standing very very accurately using cameras So it's all about sensor fusion, you know, roughly where am I? Okay, that's probably gonna come from a cell tower More accurately, where am I? Okay, that's GPS. Okay, really accurately. Where am I that's probably gonna be cameras and high high accuracy digital maps All right, and you put it all together and now you know exactly where you are You know where you're facing, you know where you're moving, you know You can do all of this and and that's the build the thing that I really enjoy about this This is just a building block. You say, oh, I've I've now got you some fundamental technology I can give you accurate position pose Direction, you know tracking etc. What are you gonna do with it as an application developer? And I think that's hugely exciting because the application developers gonna take and they're gonna do something with this that you And I had never even dreamt of but I'll bet you it's gonna be fun And this is ready now the G72 and the devices that ship in 2018 with this might be Having all these amazing things absolutely you so a lot of these computer vision applications like They call it slam simultaneous location of mapping Slam is a huge area of research that people are developing algorithms for some of which run on CPU Some of which run on GPU some of which run on specialized accelerators. It's a real sort of field of research right now You