transcribe

GTC 2013: GPUs Processing Big Data (4 of 11)

NVIDIA · 23m · transcribed 23d ago
More from NVIDIA Business
𝕏 Share ▶ YouTube 📥 PDF 🤖 .md

Transcript

0:07 look there's all all of you are working on stuff that's saving lives that's what that's what the session is going to be about today I'm going to H I'm going to Showcase a few examples of how some people are using some companies are using the GPU and and exploring the GPU to make our lives better and they're solving problems that are related to big data now to say that some of these companies are working on Big Data is an understatement they're

0:35 really you they're working on huge data they're you they're working on massive data and they have to do it not as quickly as possible because best effort is not good enough for them best effort in some of these applications basically is a fail for most of these applications and you know how consumers are they want it now they want more of it and now and so for them real time response is really vital well Twitter's it's hard to

1:07 hard to go a day it's hard to go a moment without Twitter anymore unbelievable 500 million tweets a day 500 million tweets a day and they just got started because two years ago there were only 50 million tweets a day 500 million tweets a day crowdsourcing the information that's in this body of tweets is absolutely remarkable and Smart Companies like salesforce.com are scanning this data the social landscape

1:40 continuously they're scanning the social database of every country of every social site continuously this one's interesting and so what they do is they scan the data for companies like Cisco and Dell and Gatorade who wants to know how their brand is being perceived if you launch a brand new product you want to know how your brand is perceived it's used by the Red Cross to get an early indicator of an epidemic of

2:11 people who are crying out for help using crowdsourcing information they can have instant access to information and discover early indicators of future concerns 500 million tweets now of course you're going to compare that you have to look for something within those tweets and so salesforce.com at the moment has about a million keywords that they look for a million segments what they call segments or building Expressions that they look for and they're compar comparing this

2:42 500 million tweets that are coming by all the time and continuously there's no sense saving it because you're going to get more tomorrow and it's just growing there's no way you can save this information it's like life it's just going to happen you you're going to do something about it or you're not going to do something about it so these 500 million tweets come by and they're comparing it against a million different keywords and they're searching for events indicators that

3:13 these million keywords these million things that people are looking for brands or expressions are found that search has to be done as quickly as possible or the information becomes stale while the folks at sales salesforce.com realize the exponentially growing problem that they have the number of tweets is going up the number of social sources is going up the number of customers is going up that are interested in this and the number of things that they're looking

3:42 for is going up it's a multi-dimensional expansion all probably growing exponentially and so they had the courage to try the GPU they were excited about it they ported to it and they got a 35 time speed up what used to take minutes now takes seconds what used to take minutes now takes second as a result they can really scale out this service and have a sustainable

4:14 platform to do so for some time searching for events things that are happening in society well one of the best applications on the planet is Shaz Shazam is for for for for people who no longer listen to the radio but still love music and you're no longer cool but and and so you don't know any band names the last band name you remember is

4:45 BG's the Eagles sticks that's right you know what I'm talking about I'm talking about 90% of you in the audience Foreigner when we all stopped listening to music but we still like music and we just don't have time to keep up with the new bands there's so many bands and so now you hear great music great song you just go Shazam and it tells you somehow magically what the

5:17 song is well I'm not the only person who loves it 300 million people love it 300 million users of Shazam they're adding 2 million a week that's not a business model that's an epidemic I'm sorry Jason that's a good epidemic there you are two million a week and these 300 million people are asking Shazam to identify 10 million times a

5:50 day 10 million times a day and what Shazam has to do is they have to search these 10 million queries each day these songs here take my song what's song is this and they test it and they query it and look search among 27 million other songs and they're doing this before you lose your patience and so it's got to come back relatively quickly pretty amazing stuff well to help us talk talk to us about some of

6:19 the amazing work that they do and how the GPU helps them do their work we're really really where are you Jason I think I saw you there you Jason Jason Titus come on come on stage everybody help me welcome Jason Jason good morning good to see you well you know you guys Jason doesn't do anything small before he worked on Shazam he worked on Yahoo mail and messenger now back then he only had 300

6:50 million users so he doesn't do anything small so tell tell us how did you guys discover the GPU how are you using it and and how does how does how does music search work so we had an interesting challenge that was kind of shown in that chart there and that you know as we as we get more and more users and we we went unlimited a year and a half ago and the number of searches was was growing

7:12 rapidly and we said well how can we constantly be adding new music into the system increasing with two two million new users a week and somehow meet this computational load because you we're re you have to be it's not a simple search it's actually searching in a noisy you know find the signal among the noise and do it tens of millions of times a day and so in sitting down with the with the Architects and talking what can we do to

7:35 scale this and and release all the cool features that we want to build we said you know we think we can do this with gpus now thousands of people at any given point in time have asked you to do a search and those thousands of people are asking at any second in time asking you to do a search you've got to go into a database of 27 million and growing songs how does that work I mean that's I don't

7:58 even I can't even fathom that for first it's it's worth noting that last year it was 12 million right so we we we're on the the path of growth of content is is immense and and it's also very it's not really stable if for us especially now we do we allow people to tag television you can tag any of 160 channels in the US we've done big events like the Super Bowl and that ends up being highly Spike driven so if you have

8:23 120 million people watching Super Bowl it all not 10 million linearly throughout the day it's not 10 linearly and so so the way we actually put that together is we do audio we do an analysis and come up with Fingerprints of all of those 27 million songs and and put that put together a database in memory that we can search at high speed and then all across all the mobile devices that have our software installed they when when someone wants to you know

8:48 discover and engage with media we generate a small little say 2K fingerprint and send that up to our server but that's interesting so but a song is like 3 minutes long yeah and I've used Shazam in the beginning of the song in the middle of the song at the end of the song and just the last little tale and it works yes how is it possible that you guys find the same fingerprint irrespective of where the song is tell

9:08 us the secret so we could start a Shazam so it actually that's what they want to know I'm just asking on their behalf so from the beginning we felt that we had to build something that would be essentially near millisecond accuracy like a small number of millisecond so it's actually for us when we're looking at a a 1 to 3 second sample that's actually an eternity for us you know what we want longer the times we want longer samples is because

9:32 people are talking over the music there's you're in a crowded restaurant and we need to find you know get enough of the signal coming through in that sample to actually recognize it but it ends up being I mean I can tell you before before we brought on gpus WE again and again would say hey this is great thing we want to do we want to be you know checking a whole bunch of times we want to add you know content from all

9:51 around the world all sorts of crazy features and we kept coming back to we'd love to but we don't have the capacity and so we de side we had to we had to put that to rest once and for all and build a system that could scale to having you know a billion users using us constantly every day and so now you guys have ported to the GPU how much how much capacity have you effectively Amplified

10:12 or how much throughput have you it's it's hard it's it's hard to to say it's definitely for while growing the load and increasing the increasing the amount of content like more than doubling it we also took our cost down to less than a third of what it had been wow so you know and our we see plenty of room on going to to keep that that path going so we we have I can promise anybody who's a Shazam user that there

10:35 will be very cool features coming soon that are empowered by all this and also we're hiring for folks in this room okay well if you like music they're hiring so so so so now you have 300 million users there's 6 billion phones in in use is that about right it's only the beginning for us like we we're adding 2 million new users a week we're happy about that but you know we'd like it to be 10 and there's there's 94,000

11:00 Indian movies alone yes right and right and each one of those movies has 2,000 songs within it yes and so we we I follow Twitter all the time and I find your Market opportunity is quite large that's what I'm saying we we work hard sometimes I've seen things come across in Twitter like why doesn't chazam have more Malaysian folk music and you know we you know we're we're going there but you know there's a never- ending challenge for us turns out

11:26 there's a lot a lot of music in the world lot television that people watch but that explains why it came back the other day Clueless yeah Malaysian folk music right that's terrific well congratulations Jason you guys have really done something great thank you much all right thanks for coming Jason Titus God it's so cool to work on something like that so AUD audio search music search music matching so we got we've got people matching people we got

11:56 people matching music I think you also need to match match images imagine if you were to type in F-150 for all all of the all of the press if you type in F-150 it comes up for F-150 unfortunately however if you type in Ferrari F-150 you you'll get a really really sexy car however for a lot of search it doesn't work that way it kind of works

12:26 backwards what you see is you see you see a beautiful dress in a magazine or you see somebody wearing a really cool jacket or whatever it is and you want to take a picture of it you would say I want I want a jacket like that I want a dress like that and it goes off and it finds the store it might even say what's the brand and I'll show you examples of all of the other dresses

12:53 that are kind of like that that has the characteristics of that now when when we look at it as as humans somehow we recognize what are the Salient characteristics of a particular image or a particular thing and we can identify this is a like thing now that image recognition problem is obviously complicated obviously requires computer vision algorithms and obviously valuable we would love to be able to do

13:25 things like that imagine the possibilities and so there's a there's a really cool company called cortexica and you could imagine why they call it cortexica the cerebral cortex which is used for your vision system and for understanding what you're looking at and what they're all about is to use a model of the human brain and using computer vision algorithms and a supercomputer effectively in the cloud powered by

13:55 gpus to do image recognition image processing and then image search the technique is really really fantastic and it's really at the next level now it's not matching something exactly it's not like here is a label of a wine I'm find me exactly that wine and how much it costs and what's the Parker rating where to buy it it's more sophisticated than that it's asking for what is this dress what are other things

14:26 like this thing it's also good at detecting for example things that are not straight on so for there are companies that would like to know how their brand is being used how their logo is being used in a particular photograph but it's not exactly straight on maybe it's on a on a t-shirt somewhere somebody is using the Tide logo without license and a t-shirt maybe you want to do that and and and

14:58 they they somehow could could identify that and notice that that's tied on a shirt in an in a not straight up angle and so they can recognize things that are not precise let me I think the best way to do to talk about it is to show it to you I want to you where's Mike is Mike is on the other side Mike Houston good to see J good to see yeah

15:29 so there's the latest issue of InStyle magazine let's let's show these guys what what you've been working on with the the engineers at cortexica now this is my I see my name here on U this is my copy of InStyle thankfully my wife this is the spring fashion Edition there's 594 pages of ads and I let's see here there's all kinds of great stuff let's find something oh here

15:59 we go who's that so I think that's Kate Hudson this is Kate Hudson can you guys see this so we'll show you here turn the magazine back around for me okay so now you guys that's Kate Hudson Kate Hudson and so she's wearing an an Taylor dress but and so for all of you you're probably saying well that's a cool app get me a girlfriend like that is that what you're no that's not that's the next version right well maybe they go

16:25 this version says get me a dress like that well you know there is that matching talk later you there's that matching talk you you combine this with fish.com that's you get a okay so so what we're going to do is so we're going to take a picture of Miss Hudson and we're going to sh we're going to search through 800,000 tops and blouses at eBay and in return so we're going to crop in a little bit so that we can get a so we

16:53 can match the top so it's a sort of a geometric pattern so you know let's see what it's going to return return to us so it's going to go to a GPU server it's going to generate fingerprint from the server and then it's going to be returning matches now the goal here is not to find an exact match it's to find things that are similar that you may be inspired by that are on eBay wow so

17:15 returned a bunch of geometric patterns so you got some leopard prints you got some floral patterns if you notice there's in our dress there's a little bit of a sort of flower on the side and so it sort of matched a flower pattern but you also got some of the other sort of geometric frequency patterns right so well suppose suppose you wanted you have an image of a particular particular type of pattern that you're interested in it could be a floral it

17:37 could be Asian it could be Argyle or right so so here's a bunch of patterns so well so why we try the floral pattern okay so sort of a sort of Asian style illustration of a floral pattern so the Hope here is it'll go off because it's learn databases and we'll understand that well maybe you want floral patterns or maybe you want something similar designs right so if you notice it's actually picked so if you look at this shirt right it's picked

18:01 up a full pattern from the shirt but the other fascinating thing is it's actually picked up some of the stylistic cues as well right so it's not just you know sort of Base patterns it's things that intentionally look similar so you can be inspired by and find things at eBay so again it's going through 800,000 you know tops and blouses and the turnaround time is within seconds right so you could imagine this being used for for

18:24 example art I see a piece of art I don't want exactly I don't want exactly that one because that one says on the bottom Picasso right right and so but I would like something like that you could take a picture of that you can go to eBay and it could come back and give me a whole bunch of images of art like that right so you can do things that are actually hard and Abstract right so let's look at

18:42 this one I mean how would you describe that weird Dotty speckled thing uhhuh astrophysics that's how I described it astrophysics so so let's see what what's going to return so it's a little dangerous because this one's a little bit more wild so it actually turned sort of a speckly do Matrix pattern match but again it found some polka dot patterns you know it found things that are you know somewhat related and and similar and so you can go and you know

19:06 click on these and eventually in the full app integrated with eBay you'd actually be able to just go to the buy now button and actually buy what you want so what fantastic so let me show you one more so Pinterest is really popular now sort of taken over how people sort of you know do wants and wishes and and how they like things and so people build these pages and so I actually pulled a page down on Celebrity

19:27 Fashion okay so let's take a look at at one here so this is Sarah Jessica Parker right she up in all the fashion blogs all the time so if we kind of zoom in don't get it I know so again it returns sort of floral Prints but what's fascinating is you see that red stripe that kind of goes across the top so you notice it's picked up things that instead of just being black and white also have that color

19:54 characteristic in them right so there's the sort of zebra print that has the leopard overlay on it right but there's also you know a dress that has you know red Sprites on it and if you look at this dress here oh look at that that's cool right it's picked up that base pattern so again it's about finding things that are similar they may be inspired by to buy and then taking you through to a shopping experience that

20:13 you could actually purchase that directly from somebody like eBay uhhuh wow this is going to take shopping to a whole new level all right Mike good job thank you very much what do you guys think is that amazing so you got all these gpus up in the cloud you take pictures pictures you throw added images and it comes back and show you things that are like things from other websites or other stores visual search visual search obviously is

20:40 founded on a lot of image processing technology and this is one of the areas where the GPU really rocks and rolls and we're going to see now that more and more people are experimenting with using gpus in the cloud in these massive data centers because it does image processing so much faster which is which is a relatively obvious OB observation and we have so many people uploading images now you got 6 billion phones people are taking just for just uploading to to

21:09 Facebook alone about 300 million photographs a day now today most of these websites are social and sharing photo sharing websites are really just pointing you to the photograph and often times what they have to do is they have to in what they call ingest it Lo it in change the resolution so that they could store it in a smaller form factor so that they could reduce the amount of storage but we don't have to do that we

21:37 don't have to do all that we don't have to do all the pre-computing so that so that if you were to access that that photograph again that that that the the data center could provide it as quickly as possible we could ingest resize re-encode those images so much faster now then when you add a GPU to a CPU that the number of servers these hundreds of thousands of servers that are that are running 247

22:07 uploading images uploading videos each minute there's 72 hours of new YouTube videos being loaded now in the future you want to search against that YouTube video somebody somehow used a copy Ed content in a home movie they made and they're making a fortune on that it would be good for the WebCrawler to somehow detect or extract that this

22:38 particular logo was used and they could decide how they want to use how they want to approach the the creator of the content image processing is something that the GPU does really well whether it's in computer vision image enhancement motion detection object detection the speed up by adding a GPU is really quite spectacular and so we're seeing a lot of developments in this area and we could take a lot of load out of off of the

23:04 general purpose data centers by adding gpus to them not only will we reduce the load but we will also increase its capability to do other things like video search like audio search like tagging and such

© transcribe · For agents Built with care and craft by Gokul Rajaram