 a very, very fortunate community because I mean it so happened that we are witnessing one of the biggest revolutions in the world which is called deep learning and very shortly in about I don't know couple of years or five years from now whatever we do all our lives would be touched in some way or other by deep learning and this extremely small community we are part of that revolution and we I mean we are contributing to that revolution and that's something that makes that should make all of us extremely proud about to be part of this great journey so one question that came up is from one of the audiences that what are they I think from this side of the audience so what are the applications is it restricted only to computer vision and there was some mention about the natural language processing also so the answer is no it is not deep learning as a science pants across multiple multiple sciences and segments and as I mentioned earlier it is going to touch each and every one of our lives in some way or other in a very very short while from now right I mean starting from the internet and the cloud I mean what do you buy what is your buying behavior what is that you take out from the shelf and you check the price and put it back what is the other thing that you take out from the shelf and you are satisfied with the price and you go to the billing counter and you decide to buy I mean that's that's all that is going to be analyzed by deep learning it is being analyzed by deep learning medicine biology particularly in the science of cancer reduction deep learning is is playing an extremely major big role like millions of CT scans are taken all over the world and all that is image huge image now if that is just imagine the amount of data that that that gets generated in all the CT scan centers only in Bangalore let's talk only about Bangalore if we can create a pattern out of the images that are getting generated and if some intelligence can go into it and somebody can predict in addition to the doctors if a data scientist like you you can contribute yourself and you can predict what this small aberration on the CT scan what it is likely to be isn't it great I mean this is something which is a big big field that is coming up on medicine biology media and entertainment something very similar security and defense okay there's a huge stadium 10,000 people are sitting there and how do the 10,000 people two people are not looking at the pitch they are looking somewhere else they are looking into the bag and they look very their behavior doesn't look right now how it's it's a 10,000 people stadium and the people are 10,000 people are sitting from 9 a.m. to 5 p.m. to witness the match so how does the the policeman or the security control room how do they identify two people out of the 10,000 who have a suspicious behavior and that has to happen real time I mean there is no way somebody can analyze the video after 10 days and go through 10 hours of data and look at each and every face and find that these two guys are looking funny no it cannot happen if somebody acts abnormal in a stadium cricket stadium that data has to be picked up that instance and signal has to go to the security control room and collective action should be taken to to do whatever needs to be done to if they are wrong people if they are not the right people so security and survey defense is a big big big focus area and last but not the least autonomous machines and just as I'm not going to go too deep into it so all our cars are going to be touched in some way or other deep learning and machine learning in a very short while from now and the software and all our cars are going to be software driven in addition to gasoline or diesel it's also going to be software driven I'll just deserve my comments on this because I have more detailed info on this moving forward so that sort of sums up the the the high level overview of what we're all deep learning is is is applied is going to be applied but this is only a partial list okay and this list keeps on growing all right okay so this artificial the race for the artificial intelligence is really really picking pace and it started as recent as 2010 on and summed up very well I mean he gave you a I mean I think he started in 1948 if I remember right and he he very very nicely gave you the the evolution of deep learning how what what what has happened over the past several evolution of artificial intelligence not deep learning artificial intelligence over the past I don't know 50 60 years but then last five years has been extremely interesting five years is a very very short period for any science to evolve the way it has evolved today I mean it's it's amazing I mean in 2010 that image that challenge was was announced and just five years from then in 2015 you name any major technical company it could be Microsoft it could be Amazon it could be Google Facebook IBM Nvidia Toyota Ford Tesla Audi you name any technologically advanced company from any field they are all into deep learning okay five years this has become the most important development in scientific development in the recent times and just gives a more flavor of what the numbers that we are talking about these are the numbers that are directly engaged with Nvidia I mean I just want to make it clear they we are directly touching these customers there are like tens of thousands who we don't work with and I'm sure that a big majority of you followed that category like 2013 we were directly working with 100 companies okay we know we knew them they knew us and we were working together and in 2015 it grew up to 3500 okay and this is the pace at which deep learning is going and there's an estimate that this is a 500 billion opportunity okay so I mean all I'm trying to tell is the gist of my talk so far is that this is a huge huge huge opportunity amazing science that is developing that is evolving really fast and we are all fortunate to be part of this now GPUs have sort of evolved to be a very very preferred platform when it comes to deep learning I'm sure all of you know that and so why is it so I mean is it I mean what makes GPUs so different from CPU we all know that CPUs are used for computation we all have CPUs doing fantastic work in on our laptops and desktops and and servers that sit in the data center so why GPUs are so preferred when it comes to deep learning there's a reason for that okay so let's look at the way a GPU has been architected and a CPU has been architected I mean let me clarify I mean there's nothing wrong in the way CPU has been architected that's not what I'm trying to tell but the fundamental physics the electronics that builds a CPU is different from a GPU because a CPU is is supposed to do something different than what a GPU is supposed to do how is it different so if you look at a CPU it has got a very finite number of cores I mean the diagrams show six cores and if you remember right the latest Zeon that we sell that we get from the market today has can go up to 18 cores one eight that is the number of maximum number of cores that a CPU can offer and it has got a large cache obviously now how does a GPU look like a GPU can have thousands of cores and the the latest GPU that we sell today can go up to 40996 4096 cores just imagine just compare the numbers I mean you cannot even compare how can you compare 18 with 4000 plus 4100 doesn't make sense to compare those numbers and so for any application which is highly parallel which can be broken down into multiple number of parallel threads and which can be processed to parallel e without having a sequential need obviously when you increase the number of cores it it'll run faster like I'm coming to whitefield instead of driving in on two lanes if I have 4000 lanes obviously there will be no traffic traffic bottleneck if I have 10 lanes it will obviously be much better than driving on two lanes now let's go back to deep learning we all know that as data scientists we all know that deep neural networks are highly paralyzed quotes they can be run on in parallel together and hence having a large number of cores to run your deep neural networks works much much faster than running it in a smaller number of cores and that is the fundamental difference between running your deep learning algorithms on a GPU and not having GPU to run your deep learning algorithms okay this is I mean once again I want to emphasize that I just I'm just putting forward the science in front of you why GPUs are so much preferred when it comes to deep learning now so we have a very interesting lineup when it comes to deep learning I mean any computation accelerated computing for that matter okay so there was one question there was a concern from a gentleman here whether it if it is if if GPUs are very expensive no they are not I mean they're very expensive if you want to set up a huge data center with thousands of GPUs yeah but that's not for you that's for the Googles and the Microsofts of the world I mean we are we have small startups we have huge number of startups and I know at least 10 faces in this gathering small startups who are harnessing the unreasonably effective power I mean that's what we tell when we talk about GPUs GPUs give an unreasonably efficient when it comes to deep learning and all startups can can make use of that say we have Tesla for the cloud which which obviously is meant for data centers and then then we have Titan X for the PC I mean you can have one two three PCs and have Titan X card plugged into it they're not expensive at all I mean if any of you are having a concern that moving on to a GPU based computation is an expensive solution I tell you you are wrong it is not okay please do not have that that sort of a concern in your mind I mean you can start with a couple of PCs with the Titan X card and now you all know that we have very recently we have announced the next generation which is called as the 1080 also which is even more powerful and then we have some specialized platforms for specific for automobile companies and then if you are into into creating intelligent devices like intelligent cameras or robots or chatbots etc then we have an embedded solution also I'll talk about it a bit later but please do keep in mind that GPU computation as a startup is not expensive I want to make it very very emphatic about it okay do not miss out on that all right okay so I mean as I mean there is another interesting science that is coming up I mean you should be aware of this also which is called as the hyperscale computing so we all know that high performance computing at our accelerated computing what do we do we have a huge problem a big problem and that problem is broken down into smaller number of threads and each thread is run simultaneously on in a paralyzed patient that is what we all know as the traditional accelerated high performance computing now hyperscale is another interesting development once again and the reason passed wherein it is not a huge problem that is broken down into smaller threads it is a large number of small problems okay typical and typically this is used for image processing and also for example I mean internet based applications and all the cloud based applications and all that is where the hyperscale data centers are mushrooming and keep in mind that even for that we have great solutions which are which is called as the I mean I don't want to really go into the models and all I just want to leave a message that we have solutions for all your computation needs irrespective of what your budget is and whatever your application is now I would like to introduce another very interesting offering which is called as the digits dev box this is a development box and this is specifically meant for deep learning training for creating your training networks okay now this is a very nice offering and it comes pre-packed with all the goodness or all the prerequisites of a deep learning of setting up a deep learning thing and you can choose to buy the digits dev box as it is from people or you can build your own dev box I mean all you need is a suitable server which can take up to four GPUs and there is a nice suite of software which is available which is provided free of free from nvidia just go to github and you can download it and it has all the nice software that you need in order to to create your own deep learning ecosystem and and look at the scale up I mean it is amazing for between for one GPU to four GPUs how good the scale up is when you're using such such platforms and then I mentioned about the Jetson TX one so this is our embedded platform let's say you are into creating computer algorithms for intelligent cameras or or you are working you you you are working on a project for for let's say retail stores etc etc you want a platform which which gets embedded onto your own device you are into robos you are into um drones etc that is where the Jetson which is our embedded platform plays a major role and I mean I don't want to go too deep into the speeds of the feeds but look just I'm sure as data scientists you all understand these numbers look at the numbers they are like unbelievably good I mean this this is the size of your credit card what are you seeing here the Jetson TX one it looks like your credit card and it is I would certainly say that it is a super computer by itself okay it can harness up to one teraflop of data I mean one teraflop of computation okay that's that and it has it it's a it's a it's a computer by itself it has got its own CPU memory GPU power etc etc so I want to ensure that you're all aware of such offerings from NVIDIA and and decide on what what you need to use and in April I mean many of you might have heard this term GTC that is the flagship I see a lot of heads nodding very happy about it that's the the flagship event of NVIDIA every year we have something called as the GPU technology conference GTC GPU technology conference that happens in Santa Clara and this April had been has been an exciting announcements multiple exciting announcements were made and the next generation of GPUs which is called as the Tesla P100 was announced at this and once again this is a breathtaking announcement and this is the the most advanced and most powerful GPU that has ever been announced so far and the science that has gone into this is amazing 150 billion transistors has been packaged into a credit card sized GPU which is an embedded GPU and it can go up to 5.3 teraflop of double pressure dual pressure and look at the the half precision which is it's it's more than 20 teraflops once again these numbers are huge huge huge numbers and in fact I mean I was fortunate enough to be there and when I heard for it the first time a few days before the public announcement was made I was taken aback I mean how can an embedded platform how can a small chip like thing can can do so much power and that is where I mean once again going back to the original question that we had I mean that is how GPUs are contributing so much to the to the science of deep learning and talking for particularly about the P100 because I mean why is it possible what is the science that that went into it okay this is what we call as the five miracles number one obviously is the Pascal architecture which which as I explained packaged 150 billion transistors in a small chip and then we employed multiple new methodologies we went on to the 16 nanometer FinFET technology and then the memory that we used is also a very interesting one which is called as the HBM to which is a high bandwidth memory to which which we this enables the chip to be built on a wafer on a which is which lies on a substrate that's what we call as a coarsed technology and then there is a very interesting technology which we announced which is called as the NV link I mean which we are so far we all know that GPUs the the bandwidth that is available to the GPUs is is given by the PC express bus right and whatever is the maximum bandwidth provided by the PC express bus that is what your GPU can do but what will happen if a GPU has to talk to another GPU what will happen if four GPUs have to talk to one another and what will happen if four of these GPUs and obviously they cannot only be talking to another they also have to talk to the two processors that are the CPUs that are available in the system so how is this communication why is it being restricted to the bandwidth of the PC express that is what is answered by this great technology which is called as the NV link technology and of course the new artificial intelligence algorithms I mean we are no longer a GPU only company we are a platform company and we the the importance that we give to generate to to the R&D that is the time and the resources that we send spend in creating great GPUs equal amount of R&D also goes into the software and we are slowly evolving into a we have evolved into a great software company also and all these the goodness that I talked that I talked about for GPUs they all come bundled with amazing improvement when it comes to the algorithms I'll be talking about it a bit later now oh I think this sort of repetition and last but not the least there is also a page migration engine this is a new technology that we have introduced on the P100 GPUs wherein you have a unified memory that is available for the CPU and the GPU that the system memory and the GPU memory are pulled together as a common pool and that huge pool is available for your computation okay no longer only the GPU memory is available for running your computation on the GPU you also can tap into the huge system memory or the RAM what we commonly call as and that is possible due to this technology new technology called as the page migration engine and so what exactly how exactly does it translate to when it comes to performance huge huge leaps in everything okay so this Tesla M40 whatever you are seeing here this was announced about I think about six six months ago six to seven months ago is when it was announced and obviously this the P100 is what we talked about and it can give three times computation and three times GPU memory bandwidth once again thanks to NV link and also five times GPU to GPU bandwidth okay now this is the type of improvement that we are talking about and this is a breathtaking pro product okay and it has already been announced and a lot of experimentation is happening it will soon be very commercially available in in I think about three to four months from now Q4 of this year is when it will commercially be available to all of you now in addition to announcing about the P100 we also came out with an announcement which is called as the NVIDIA DGX1 I'm sure if at all some of you have been following on GTC you might have some of you might have heard about it and this is the world's first deep learning supercomputer okay the world's first deep learning supercomputer it is a system by itself it's it's a full-fledged supercomputer and it's just a rackable one 3R form factor and it is specifically engineered for deep learning and it can deliver this one box can deliver 170 teraflops of half-pressure operations okay 170 teraflops I'm sure that people who are associated to the accelerator computing world you would you would you you would understand how big the type of computation is when I say 170 teraflops it's it's huge it's an unimaginably huge and it it encompasses eight tesla P100s and a whole lot of ecosystem that goes along with it and very importantly it accelerates all the major artificial frameworks that you are aware of okay so I mean this is just a block diagram which shows you what goes into a DGX1 all right so what does it translate it into I mean it's great 170 teraflops but is the physical way of me understanding it what this 170 teraflops means is to put it in very simple terms it can give you the compute power of about 250 servers that you are currently having in in your data center okay this one 3u super computer delivers the same performance of 250 servers okay without GPUs that are there in your data center so that is the equivalence of the power and so once again I would like to I mean remind you that if you are in deep learning please please do try out GPU otherwise you are losing out quite a bit and once again I mean this is GDC 2015 is when we announced the previous generation Maxwell architecture and what took 25 hours has gone down to two hours in just one year okay now I want to remind here that all these improvements are not only because of the silicon not only because of the hardware it also is because of the the amount of software the R&D that we that goes into the software that are available with GPUs okay so it's an end-to-end product family whatever we have seen so far for hyperscale computing for cloud based data center computing and for your embedded systems and dgx1 last but not the least which is the supercomputer specifically meant for deep learning training okay and this enables deep learning everywhere okay it could be a small startup using one workstation with titanics plugged into it or it could be a robos with the embedded platforms or it could be a data center with the Tesla or the car which is running around which has a specialized platform which is called as the drive px deep learning is everywhere that sort of sums up the the hardware port of it that I wanted to to tell you now let's move on to software and this is even more exciting than the hardware that we have discussed so far okay so when it comes to the deep learning SDK it is a continual improvement we keep on improving it again and again and again and and you can see some code from from the architect of tiano there the world is amazed at the type of improvement that we bring to the software that goes hand-in-hand with the hardware also so let me this is sort of a block diagram which gives you the software ecosystem that we provide with the GPUs for deep learning obviously the top applications today are computer vision speech and audio and natural language processing these are the top ones okay the maximum used ones the other others that I mentioned earlier internet and cloud and big data and medical imaging yes they are certainly there but these are the top ones and then you have a big family of deep learning frameworks that are available I mean I'm sure many of you are available are very familiar with these names that you have cafe you know your CNTK from Microsoft you have tensorflow from Google all these are very very popular they are the popularity keeps on growing and that their capabilities are also increasing and then you have so this is the ecosystem that is available and NVIDIA SDK provides multiple tools I'll be very briefly touching upon each of these tools and as data scientists and deep learning algorithm people you will be able to relate to the work that that I will be explaining we it spans across all these segments whatever could be your application or whatever framework that are using there are there are the best tools available to utilize the hardness of to utilize the power of GPUs okay and you name any deep learning framework okay they are all accelerated by GPUs and the SDKs that we provide all right so once again in GTC we announced something called us a NVIDIA compute works with this is basically the compute works is just a bundle okay and though there are four major advancements when it comes to NVIDIA computer works which is CUDA and Qoody NN and NVGraph and index okay I'll be touching upon each one of them very shortly and the big announcements or the improvements that we have made is in terms of digits I mean some of you might remember I mentioned about the digits development box which is a deep learning training system okay having capable of taking four GPUs and digits is the software that goes into the deep learning training system and then you have something called as a GAE I'll be talking about a bit later which is graphics inference engine okay graphics inference engine which is specifically meant not for inferencing okay we all know that we have been great when it comes to training how about inferencing so that is what is addressed by GAE graphics inference engine and you have an improvement on the Qoody NN which is a derivative of CUDA all right now let's start from the basics let's start from the basics I'm sure CUDA is very commonly known term to most of you here I mean what is CUDA CUDA is a parallel programming platform that is offered by NVIDIA and that is the software platform that is used as the base to build everything I mean be it the conventional accelerator computing or deep learning algorithms whatever it is it is all based on CUDA which is our parallel programming platform now CUDA was very generic so when deep learning as a science was evolving what we did is we customized CUDA and created multiple forms of CUDA which are tailor-made for specific applications that is what we are going to talk about now and CUDA as such the basic CUDA as such is undergoing tremendous improvement in performance and as I can see within a span of two years what we had in 13 and what we are having I mean we had in 2015 the performance of CUDA went up to five and a half times okay now all right okay now I mentioned that CUDA and then which is which is a very very specialized form of CUDA which is which expands to CUDA for deep neural networks okay so please make use of this and the best part is this CUDA and then is available free I mean so long as you are having your GPUs it is it is it is free it is not open source it is free all you have to do is you go to the developer site website in NVIDIA and you can make use of it and it is going to give you great performance when it comes to any framework that you might be using all right and there are multiple specialized as I mentioned specialized forms of CUDA something called as a CUBLAS which is your if you're using it for your linear algebra programming or you have CUSPARS for your SPARS matrix solvers etc and whatever and you have another very specialized tool which is called as NCCL which is a parallel programming tool which which enables scale up of your performance across GPUs when it is inside the one node or across multiple nodes also there are so many nice tools which are available which is what we call as the NVIDIA software development platform and if we wanted to ensure that you're all aware of this and please explore this and start using this all right okay now talking about GPU inference engine we all know that GPUs have conventionally been very good at training now with the GPU inference engine the GIE inferencing also has become fantastic and very advanced inferencing can happen with the help of this GIE once again this is part of the the P100 announcement that we made and watch out for this and start using this and you'll be amazed by the performance of inferencing that you are getting here and last but not the least graph analytics if you are into big data if you are if you are into finding out relationships between millions of instances how each instance interacts with the other then you have a great tool which is called as the NVGraph now I would like to once again emphasize here that all these are based on CUDA they all are formed upon one one particular software platform which spans across this so you can develop today in Titan X okay and you can the same algorithm with some recampillation can be ported on to your data center and it can be put it on to your embedded platform or it can be put it on to your dgx1 box at some later point in time you start low you start small you start on your PC or a workstation and the same algorithm can be put it on to into multiple platforms at the right time arises and the last topic that I am going to cover in a couple of minutes is what is the amazing year for self-driving cars I'm sure all of us have heard this this terminology Google car and and a lot of research that that goes into cars all and NVIDIA is playing an extremely big role when it comes to the autonomous the automotive industry also and all the major automotive companies are getting into this and they're all moving into the Bay Area to set up R&D facilities that can help in automotive in self-driving okay now obviously self-driving as many of you know is is is certainly a deep learning problem wherein the computer vision will have to identify passenger identify people who are crossing the road or it has to read the signs on the road or it has to take note of the policeman signs and act accordingly so obviously it is a computer vision problem and it's a deep learning problem and so that the new artificial intelligence driven car this is how it works so you have the NVIDIA DJX one which is used for your training which which goes into the car and there is a special platform which is called as the dry NVIDIA Drive PX which sits on the car and deep learning happens on the car I mean it it's not that it goes to the cloud and comes back the deep learning happens inside your car and the right decisions are taken and your car is driven by using localization and mapping okay so self-driving loops this is what happens doing a typical deep learning problem when a autonomous car is driven so a mapping of the terrain is happen is happening and then a localization to understand where you are I mean on and touched upon it I mean look at this call I need to understand where you are you're you're existing in a specific defined space that has to be done on on the road also and then the computer sees and then the right decisions are taken to ensure that the car is driven without the help of a driver and that is what we are talking about in autonomous driving system now it is not only autonomous driving system I mean many of you might think that how is autonomous driving system applicable in India it's going to take a long time it's also the the human machine interaction the HMI that happens is going to change a lot and all our cars are going to be more software driven in the days to come than anything else so we are going to have the next generation user experience and this is how all our cars are going to look like in a very very short while from now wherein you have a very nice user experience and you have all your cars are going to you will be able to talk to the car the car will understand what you're talking natural language processing and it'll react to your commands you can just sit inside the car and say take me to the airport and it'll be taken you'll be taken to the airport and these are all I mean happening I mean these are these are great sciences which are happening which are evolving and this is I mean I don't know how many of you are aware of this is amazing so in by the end of this year we are going to have a race which is called as the Robo race okay and all these cars are self driven okay there are 20 cars which will be self driven 10 teams 20 cars and one mandate is that all the 20 cars must be identical from the hardware point of view okay only the algorithm has to be different and who writes the best algorithm his car his or her car comes first and here in wins the race and needless to mention all these 20 cars the drive PX which is the Nvidia platform is is the brain that drives the car so great things we have talked about have a touch have I explained everything about about GP competition what do you think not sorry questions yes and this my last line not really I have not explained everything about deep you can be combination I've just touched the tip of the iceberg if you want to know more about it if you want to learn more about it I invite all of you to the first ever GTC the GP technology conference we are very happy to announce that it will happen in India not in Santa Clara in India in Mumbai on December 6th we really want to see all of you there you can present a technical paper you can present a poster or you can simply add and learn like what you're doing today I mean most welcome and last but not the least they we also are conducting an emerging company summit okay which is part as part of the GTC and present your papers take part in the competition and the awards are amazing I mean whatever particularly if you're a startup you have to take part in ECS in emerging company summit keep watching out our web space and you will get more and more information on this all right that is what I wanted to to cover and once again Mumbai December 6th we really look forward to seeing you there in GTC India okay thank you questions we have questions questions and then you can catch on the offline later hello yes hello is NVIDIA selling Digits DevBox in India I'm sorry is NVIDIA selling Digits DevBox in India as of now no but you can help you can build your own Digits DevBox which is exactly identical to what is commercially available and we can help you in building it hello yeah hi this side in last GTC 2016 we saw that NVIDIA was actually collaborating with many universities could you speak up I'm not able to hear you hello in last GTC 2016 we saw NVIDIA teaming up with many universities they were actually collaborating with US universities and there they were giving some hardware especially DGX-1 to the universities for their research work can we see similar things happening in India yes the answer is yes we certainly have very structured university relationship program one of my colleagues is managing it the answer is yes we have big plans for academic institutions in India contact us and we will give you more details yeah can we know about the contact details yes I'll give it to you just that when we have full oh you can write to me it is available on the screen just no down the email and the phone number you can write to me I'll pass you on to one of my colleagues who handles the university relationship program thank you so much yeah I think there's just time for this last question so yes yeah I'm here hi yes I'm just wondering how does your hardware compare with our human hardware of brain how does our hardware compare with what the capacity of human brain in terms of neurons etc okay how does it compare the capacity of the brain okay okay sure thanks I mean so we have multiple levels of accuracy that that that are used to to measure the the performance of a deep learning algorithm and most of the times we find that I mean I think Anand also mentioned about it recently most of the times we find that the accuracy that gets thrown out of a deep learning algorithm is even better than that of a human brain okay and I'll bring a very interesting perspective to this so one common concern about autonomous driving is that how safe it is but the studies show that when the world becomes fully autonomous in terms of driving autonomous driving will be safer than humans driving around okay yes I think should take the questions offline everybody's proceeding towards tea please be back in the auditorium by 1140 because we will start the next talk at 1145