Jensen Huang, Founder and CEO of NVIDIA with Ali Ghodsi, Co-founder and CEO of Databricks

so I am super super excited to introduce our uh next guest uh who actually is a man who does not need any introduction so I want to welcome the world's one and only Rockstar CEO Nvidia Jensen Wang to [Music] [Applause] [Music] [Applause] stage thank you so much awesome thank you for coming so uh Matt I just want to start uh you know just looking at nvidia's amazing performance you know three3 trillion dollars like did you did you imagine it would be this way say five years ago that the world would unfold this way sure from the very beginning it's so awesome to see any advice for a you know fellow CEO how do we get there what whatever you do don't build GPU okay all right let me tell the team we need to back out awesome man uh so we spent a lot of time this morning uh talking about data intelligence yeah which uh by by what we mean Enterprises you know they have all this proprietary data training AI models that's customized on their data that they have how important is that is that something you see you know uh you know is that something that we need to invest more in what what are you hearing well every company's business data is their gold mine and there's every company is sitting on gold mines if you have if you have a uh a flywheel of services or products customers enjoying those services and products giving you feedback you've been collecting data for a long period of time it could be customer related it could be Market related it could be supply chain related um all of us have flywheels of data that we've been collecting for a long time we're sitting on M of it but the fact of the matter is none of us have really been able to extract Insight or even more importantly distill intelligence out of it until now and so we're pretty fired up I know we are yeah and we're using it in our chip design we're using it in uh our bugs database uh we're using it uh in creating new products and services and we're using it in our supply chain and you know for the very first time you we now have a business we now have an Engineering Process that starts with data processing and refinement and learning models and then deploying the models and connecting that flywheel collecting more data wow isn't that right yeah and so we're doing that in our company and is make making it possible for us to frankly be one of the largest small companies in the world and and the reason for that of course is because we have so many AIS in the company helping us out doing amazing things and I think every company is like this you know and so so I think this is just an extraordinary time and it starts with data it starts with data bricks that's awesome thank you so much um curious you know there's this whole debate Brewing uh closed models versus open source models you know are open source going to catch up uh is you know are both going to exist is it going to eventually just be dominated by you know a one giant close Source model what are you seeing what are you thinking about the whole open source ecos system and how important has it been for sort of development of llms and how important is going to be going forward well we need we need Frontier models we need amazing Frontier models of course the work that uh open AI is doing the work that Google's doing um uh really really important and pushing the Frontiers and and helping us helping us discover what's possible uh but if you were to to look at this year probably the most important events this year were related to open source uh llama 2 now llama 3 uh mraw uh the work that you guys did data bricks um uh dbrx do I have to say dbrx dbrx right dbrx dbrx dbrx I think really really cool stuff and the reason why it's so really cool is because it activated every sing single Enterprise company it made it possible for every company to be an AI company isn't that right you're seeing this yourself and we're seeing this all over the place uh we recently turned llama 3 into a uh fully containerized uh inference microservice and it's available for downloads you can go to hugging face and go to of course uh data bricks and it's now it's now being inter ated into several hundred companies around the world and so that tells you something about how open source has activated every company to be an AI company we're using open source models all over our company and and we create some proprietary ones uh we uh uh uh fine-tune um uh open- Source ones uh train them for our data and for our skills and so so I think without open source it wouldn't have activated this entire Global movement for every company to be an AI company I think it's a huge deal yeah that's super super awesome um so both are going to be around and we need both like open and close uh and this is is this the Nim framework you're talking about Nims uh of how you do the survey yeah we call them Nims yeah yeah yeah we're super excited you know I'm super excited to announce here that we're going to put dbrx inside Nims and we're going to serve it on data breaks and actually any new models that we develop in the future so we're super excited about NS yeah it's it's actually quite quite an amazing thing in order to in order to create one of these one of these endpoints these apis these large language model apis the stack is really complicated you know these are giant models uh even though even though you know they they seem small these days they're still computationally really large and the Computing stack is really complicated there are hundreds of dependencies necessary to create one of these endpoints yeah and so uh we created this thing called the Nvidia inference microservice where we package up all of the dependencies we optimize all of it we have a factory in the company with all these Engineers working who are expert in doing this and we package it up into a microservice and you could enjoying it enjoy it at at data bricks uh you could U download it and take it with you uh you can fine-tune it with uh microservices that we call Nemo and uh uh use it anywhere you like it runs in every single Cloud runs on Prem uh you know you can enjoy it everywhere that's amazing thing yeah yeah and it's awesome you can even run it on Prem I mean that's that's you know it's not you don't have to be on the cloud um that's super awesome okay so when we talk to customers we're hearing that you know they have to develop this uh sort of expertise in house to customize models to gain Advantage what are your thoughts on that well I think I think in the future um look what's happening in the world today is that that we figured out a way to tokenize almost any information almost any data and we can extract structure understand uh learn its representation understand the meaning of that information um of almost any kind it could be of course sound speech words language images videos uh it could be chemicals and proteins it could even be uh robotics articulation and manipulation it could be steering wheel articulation driving uh we can tokenize almost anything and because these these uh these uh cloud data centers um are really producing tokens we're manufacturing something that is quite unique for the very first time you have this instrument called these AI supercomputers that we build uh it's producing tokens yeah generating tokens in essentially a uh Factory that's designed for that one job and this ability for us to manufacturer intelligence at scale is pretty new and that's one of the reasons why I'm almost certain now as we're building these AI factories everywhere uh for all these different industries that we're in the beginning of a new Industrial Revolution instead of generating electricity uh we're generating intelligence every company of course at its foundation is about domain specific intelligence yeah uh very few companies on the planet knows more about data and data processing and Ai and uh the infrastructure necessary to do all that then data breaks um we are quite specialized in the work that we do and we're uh at this Foundation all about that domain specific intelligence every company is it could be Financial Services could be Healthcare whatnot and so at the end of the day every one of us will become intelligence manufacturers and if you're going to be intelligence manufacturers today you have HR in the future you're going to have you know HR for AI and we call them AI factories so every single company will have to do that we are doing that you're doing that um uh we see companies large and small doing that and so in the future uh 100% of us will do that uh you start with of course your domain specific data it's sitting in data bricks somewhere you're going to process uh that data and refine and extract intelligence out of it you're going to turn put it into a flywheel you're going to have an AI Factory all of us will yeah this so so awesome I totally 100% believe in this and one thing we're excited about is you know so we do a lot of data processing and data processing it's like massive amounts I think we process about four exabytes every day you know 4,000 terabytes every day in data breaks and you know it is the single largest Computing Demand on the planet today processing data yeah right every single company does it yeah exactly and you know it's actually highly paralyzable you know we do the same operations again and again again so I'm really really really excited to partner together uh to really bring that kind of GPU acceleration to data processing so we can do the same Revolution that AI models have seen on the core data processing so we're super excited to partner with you on using GPU acceleration for our Photon engine to be able to really kind of enter this new era of also applying gpus to core data processing right these massive workflows that today have to run on CPUs getting them also run on Nvidia gpus we're very excited about that yeah this is a big by the way this is a big announcement yeah the two the two most important Trends in Computing today is accelerated Computing yeah and generative AI yeah Nvidia and data bricks are going to partner to combine our skills in these areas to bring it to all of you and [Applause] yep and this this work in accelerating data processing as you know it's it's highly paralyzable yeah but it's really Arcane it's really complicated and the reason for that is just there's so many data formats there's so many different ways to group and join and you know just wrangling data is a really complicated Suite of libraries spark is a super complicated Suite of libraries and it's taking us 5 years working around the clock to finally have a suite of libr liaries that can now accelerate Photon and this is such a big deal we've been working on this for a long time for many years so now we're going to accelerate Photon and make it possible for all of you to Wrangle data process your data a lot faster a lot more coste effectively and very importantly consume a lot less energy yeah makes a lot of sense a huge deal y it makes a lot of sense right because it's uh in the end of the day even though it's very complicated and you know it has lot of corner cases it is highly paralyzable and uh you know it is specialized still it's not you don't really need generic compute for that right it's it's we want to do it's like you know same thing again and again and again on xaby to data right it's not we're not doing xaby to data that's completely unique so I'm very very excited about this and I think it's really has the ability to revolutionize and really bring faster you know performance lower the cost and just you know uh it's going to be amazing yeah look look what happened when we're able to process enormous amounts of data so quickly it made a possible for researchers to one day wake up and say guess what let's just go get all of the data on the internet and train a giant model because it doesn't take that long without acceleration without accelerated Computing nobody would have ever conceived of doing that it would have been way too expensive would have taken too much time but now you know it's kind of a mundane thing to do so you know we're going to be able to process exabytes and exobytes of data so much more cost effectively and so much much more efficiently from a Time perspective imagine all of the ideas that you're going to have it's you know it's going to be hey let's just take all of the data of our company and we're going to train our super AI you're going to do it yeah the day is going to come yeah I mean it was a Sci-Fi idea right to take the whole internet nobody thought you could do it we needed the hardware to get there the infrastructure to be there so we could specialize it and now you know everybody's doing it um so I want to switch gears God I love myself just kidding just kidding we love you too uh so um I want to switch gears um so you know this generative AI boom has been amazing um you know but the early days you know most Enterprises started with chatbots let's build our own chatbot you know customize it on our data and so on but now we're seeing people Branch out to more and more sophisticated use cases yeah what new applications in AI are you the most excited about going forward um the number one most impactful will probably be customer service for all of the Enterprises that are here customer service you know represent probably several trillion dollars worth of expenses and every company has it every company has it every single industry has it every company has it um and and um the important thing about the chatbot the customer service is is partly about the fact that you could automate but it's mostly about the data flywheel yeah you want to you want to capture the the the conversation you want to capture the engagement in your data flywheel it's going to create more data of course we're probably you know right now we're seeing data expanding about growing about 10x every 5 years wow I would not be surprised to see data growing 100x every 5 years because of customer service and so we're going to we're going to connect everything into a flywheel it's going to collect more data capture more insight we're going to extract better intelligence out of it which will provide better service maybe it's even more predictive in the sense that proactive in the sense that uh before a problem even arises you reach out to the customer and say you know this thing is about to expire or um we notice that you're still using this version or whatever it is and you reach out to the customer and proactively solve a problem just like preemptive maintenance we're going to have Pro proactive customer support which is going to create more data we're going to write that flywheel and so I think I think customer service is probably going to be the most profoundly uh supercharging capability for c for for most companies because of that because of the data it's going to collect but we've tokenized everything you know the I'm excited about the fact that we're generating chemicals we're generating proteins um uh we're uh carbon capture materials uh carbon capture uh enzymes uh in incredible batteries that are being being uh designed uh and so we're generating physics uh phys physical AI uh and recently we um made it possible to do Regional weather prediction down to a couple kilometers now it would have taken a supercomputer about 10,000 times more capability to be able to predict weather down to a kilometer and now we're doing we're using generative AI to do that wow um and and so as a result uh Logistics will be enhanced insurance will be enhanced uh of course uh keeping people out of Harm's Way will be enhanced um and so so physical things uh biological things uh of course uh you know generative AI for 3D Graphics digital twins uh creating uh creating virtual worlds for video games I mean generative AI is just uh everywhere every single industry if your industry is not involved in generative AI is just because you you haven't been paying attention it's everywhere yeah yeah I totally believe it you know we're going to see there's no no area where we're not going to see applications of this makes a lot of sense it's so exciting um you know these New Frontiers are super exciting and there's huge needs for data you know AI what's your thoughts on how we can help Enterprises make AI That's more sustainable well um there's a lot of sustainability has a has a lot of a lot of different perspective one one of the sustainability uh has to do with energy yeah and and um remember AI doesn't care where it went to school you we could we don't need to put AI training uh data centers uh near population where the energy grid is challenged already we could put it somewhere where it's not challenging and so uh you know that the world has Earth has a lot more energy uh it's just in the wrong places and so I think for the very first time we can go capture that excess energy press it into an AI model and bring these you know AI models back to society where we where we could use it uh that's one one one major thought and another is um I remember that AI is not about training AI is about inference yeah and it's about the generative capabilities of the AI you're training the model so that you could use it uh and when you think about the longitudinal benefit of AI and I just gave you the example of of uh predicting weather using AI instead of using supercomputers we understand basically the laws of physics that's involved in weather prediction we don't need to simulate it from first principles every single time we got to generate it using Ai and by generating it using AI not only do we reduce the amount of time that it takes improve the resolution that we can generate for but also the amount of energy by thousands of times not tens of you know not percentages thousands of X factors well by doing that we're doing the same thing by designing chips that you're using in cell phones um say you know you you train the model ones design better chips with those models as a result you save energy for everybody involved I'm just when you think about the longitudinal benefit of AI I'm fairly certain that it will demonstrate the amount of energy that's saved and then one last thought about about generative AI you know that today's why is such a big deal from a computer science perspective today's Computing experience is retrieval-based largely you know we touch the phone and even though uh when we when we use our phone we think it uses very little energy every single time you touch it it goes off and sends rest activates rest apis all over the world retrieves information the internets on you know lit up uh brings back a little bit of information for you from all these different data centers assembles it based on a recommender system presents it to you well in the future it's going to be more contextual more generative right there on the device running a Model A small language model the amount of internet traffic will be dramatically reduced and it'll be much more generative with some retrieval to augment right and so so the the balance of computation will be be dramatically shifted towards immediate immediate generation well this is very you know it's going to save a ton of energy and it's very sensible and the reason for that is this imagine every single question that Ali asked me I got to race back to my office go get some files bring it back and present it to him let him decide which piece of that information he wants to extract for himself instead I'm generating everything you know from about 25 watts right now as we speak yeah right and so the amount of energy that we save is going to be extraordinary and the Computing model is going to transform completely and so this way of this way of uh of computing is going to save tons of energy U of course we're going to get our answers a lot more efficiently instead of us combing through stuff but then we'll have even more questions right we'll have more questions which is really in in fact that's the big idea the big idea about the future as working with AIS is prompting we're going to have so many more interesting questions because we're going to get a lot of answers very quickly yeah so this is a very big deal yeah very exciting Future Okay my final question to you how do we help customers you know organizations here get started today what's the best way well you know I I told you before that I thought the pivot um of data bricks expanding from data processing to data governance and uh store uh and then extending it into all the way uh longitudinally all the way to extracting uh intelligence out of that data I think that that was completely genius and and uh I forget her her name but I thought cookie lady did an incredible job Casey what's that Casey okay don't steal her [Applause] please I I thought she did an amazing job I was enjoying I we were in we were backstage and and everybody wanted to talk but I just want to watch her give her demo I I thought I thought the platform uh is incredible and you've made it easy for people to uh manage their data extract information process that data Wrangle that data you know wrangling data is still a very big part of training the model people talk about training the model but long before you train the model you got to go figure out what data right it's about data quality it's about data format it's about data data preparation and so so I I think I think the way you start is is uh come to data breaks and uh use the use the data breakes uh data intelligence platform am I right yeah [Applause] absolutely who who wouldn't call their platform dip that such a good idea you know dip data breaks dip sounds good I like it it's almost as good as Nims all right all right there you go there go well you can do both together right yeah you don't have to pick go get yourself a Nim on dip yeah I agree why not that that's the way to do it that's that's you can dip I would absolutely start whatever you do just start whatever you do start you have to engage you have to engage this incredibly fast moving train remember generative AI is growing exponentially you don't want to wait observe an exponential Trend because in a couple of years you'll be so far behind it's incredible just get on the train enjoy the train as it's getting faster and faster exponentially learn along the way and so you know this is this is one of those things you can't learn by watching you don't want to learn by reading about it you just learn by doing yeah and which is the way we're doing it and so just get engaged all right that's great advice Jensen it's been an amazing decade thank you for everything we've been great Partners looking forward to our next decade together data bricks all right

Channel: Databricks

Views: 35,814

Keywords: Databricks

Id: SAsoWmMhX3Q

Length: 25min 2sec (1502 seconds)

Published: Fri Jun 14 2024