# Microsofts New AI Chip 'Athena' SHOCKS The Entire Industry ## Метаданные - **Канал:** TheAIGRID - **YouTube:** https://www.youtube.com/watch?v=JJRDqSiHRrE - **Дата:** 29.04.2023 - **Длительность:** 9:35 - **Просмотры:** 41,934 ## Описание Microsofts New Chip 'Athena' SHOCKS The Entire Industry Welcome to our channel where we bring you the latest breakthroughs in AI. From deep learning to robotics, we cover it all. Our videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on our latest videos. Was there anything we missed? (For Business Enquiries) contact@theaigrid.com #LLM #Largelanguagemodel #chatgpt #AI #ArtificialIntelligence #MachineLearning #DeepLearning #NeuralNetworks #Robotics #DataScience #IntelligentSystems #Automation #TechInnovation ## Содержание ### [0:00](https://www.youtube.com/watch?v=JJRDqSiHRrE) Intro so the AI race is heating up once again and not in the way that you might think this is Microsoft's new game plan they are working on their own chips that are going to rival and videos now I gotta be honest if you don't know how far ahead Nvidia are you should honestly take a look now before I get into showing you all the details of nvidia's chips Microsoft's new secret chip and later on in the video even Google's new secret chip which they claim is faster than nvidia's I need to tell you all why these chips are so important to a chat gbt and air software so essentially you all know exactly what chat GPT is and how these AR softwares are used in our daily lives now the thing is AI is ### [0:37](https://www.youtube.com/watch?v=JJRDqSiHRrE&t=37s) What is AI actually learned through a process called Deep learning now deep learning involves training AI networks with large data sets to learn from and examples to make predictions or decisions so essentially you know a chat gbt how people feed at large amounts of data and in mid-journey how they fit at loads of pictures that's how these programs are run now of course these processes are computationally intensive require sharing massive amounts of data and complex mathematical calculations now Nvidia makes ships that are optimized for this type of processing with features like tensor cores that are designed to accelerate Matrix math operations which are Central to many deep learning algorithms so the chips that Nvidia make are also highly scalable which make it suitable for large machine learning applications which essentially means that these are the kind of chips that we need to power the next generation of AI now what's also cool about these chips is that they support high bandwidth networking and advanced memory Technologies which allows for faster data transfer and more efficient use of memory so essentially these chips are literally an integral part of the next World's Industrial Revolution and you have to understand that without these chips it's going to be very hard to progress in AI now here's the thing okay Nvidia make these chips and they're pretty much the best chips that money can buy they recently had the a100 which was a decent chip but even recently Nvidia upped the bar once again with the h100 chip that you're seeing on screen right now and this one is super powerful that is completely optimized for AI now here's the problem okay now Nvidia is just one company that makes these chips and you can see right here on eBay that these chips are very expensive and highly in demand now what's crazier is that you can also see right here that the cost of training gbg3 was around 4 million dollars which is honestly insane and what's even worse is that openai spends approximately three million dollars per month to maintain chat gbt and that's just based on the many courses that come with running these highly expensive gpus now you have to understand that if these companies don't manage to create their own chips that they can make in-house they're going to be 100 reliant on Nvidia and if Microsoft is reliant on Nvidia then they're simply not going to be able to compete because of course that is going to be a hefty cost that they have to keep spending on and of course Nvidia is going to be the only one who's raking in all these profits right now the AI race is heating up for more than ever and everybody knows that large language models are going to be a thing in the future where the company who manages to master these chips is essentially going to be powering pretty ### [3:10](https://www.youtube.com/watch?v=JJRDqSiHRrE&t=190s) Microsofts Athena much the entire world to large language models because they won't have a choice so that's why companies like Microsoft and companies like Google are racing to make their own chips this is because if they then they're less reliant on the Behemoth that is NVIDIA and of course we do know that these companies want to develop these things as quickly as possible as large language models get more and more scalable and more effective now that you know exactly what these AI chips are for let's talk about Microsoft's Athena so allegedly 300 staff are actually working on these ships with the sources saying that we could see wider use by both Microsoft and openai sometime next year although they are unsure whether or not they actually want this chip to be available to cloud computing customers however this article goes on to state that openai has actually used Microsoft super Computing systems to train their large language models that power chair gbt however these computers actually run on nvidia's chips which is of course a problem that Microsoft is quickly trying to solve as these costs get more and more expensive now if you don't realize how expensive these models are remember every time you go on to chat GPT and even if you are a plus subscriber you try and access gbt4 you see this message right here it says gbt4 currently has a message cap of 25 messages every three hours I'm pretty sure that's largely due to the fact that gpt4 is much more expensive to run than gbt 3. 5 which means that they definitely need to make this much more effective much more quicker now remember we previously talked about earlier in the video how Microsoft isn't the only company who are developing their chips we also have Google teasing its tpu4 okay and this is an insane chip because they actually claimed that it was faster than nvidia's model the a100 which is also very very quick which is specialized in deep learning and AI capabilities so in this article it says Google claims it's TPU V4 outperforms nvidia's a100 so a new scientific paper from Google details the performance of its Cloud TPU super Computing platform claiming it provides extra skill performance on machine learning with boosted efficiency the authors of this paper claim that the tpu-4 is 1. 2 to 1. 7 times faster and uses 1. 3 to 1. 9 times less power than the Nvidia a100 in a similar sized systems so they note that of course by this being a very impressive metric they haven't actually compared this to nvidia's h100 GPU due to their limited availability and their architecture now of course as we stated previously in the video These are extremely rare graphics card because they are of course also very expensive now you might be wondering so just how powerful are nvidia's h100 gpus compared to Google's cloud CPU V4 supercomputing platform where they're trying to build this machine learning infrastructure well just take a look because honestly Nvidia seemed to constantly push the barrier and seem to be Miles Ahead from the competition so take a look at this okay Hopper architecture accelerates AI okay so basically nvidia's new h100 GPU is essentially nine times faster up to 30 times faster on training AI this is absolutely incredible now the reason I say is because you've got to remember okay this is faster than the previous GPU but remember Google was recently boasting about how their own gpus were faster than nvidia's a100s but Nvidia just released the h100 which is actually nine times and 30 times faster in certain aspects which I think it's important to note that the rate in which AI is developing now many people are talking about large language models which of course are very interesting and very effective as well what they do but what you're not paying attention to is of course the hardware side in which these companies are racing forward to develop these chips that are literally going to power the next Industrial Revolution which many are referring to as Ai and I do think that it is important to note the role that these gpus and that these chips will play in advancing this and making everything much more quickly because eventually we do want to have this in real time I mean chat TBT is great but if we could have this information instantly available to us in real time that would be far more effective now there's also something that is incredible which is a breaking area of technology which I do just want to quickly touch on so this was a video that came up in my recommended feed because I've been consuming a lot of content on AI now this video is very interesting because it talks about brain like Computing and spiking neural networks now let me explain why this is literally the next level in Ai and once this technology is mastered we're about to see a major break through the video actually talks about neomorphic Computing and essentially it's basically Computing that models the way the human brain works so it involves creating artificial neural networks that can process information similar to the way that the brain does so basically the brain as you know is made up of neurons which are cells that are responsible for transmitting information when a neuron receives a signal it sends out its own signal to other neurons and this process of signaling between neurons is what allows the brain to process and interpret information now neuromorphic Computing works by creating artificial neurons and then connecting them in a network so these artificial neurons are designed to behave like real neurons so they can receive and send out signals just like the neurons in the brain so by connecting these artificial neurons we then can create an artificial neural network that can process information in the same way that the brain does so the benefits of this okay are that this is basically like a human brain and the reason people are investing and companies are developing these applications such as Intel is because us these are what will power the next AI transposition so it's still a relatively new field but this video dives into a lot of stuff like 2D materials which are essentially materials that are so thin that they literally call them 2D materials because they're literally as thin as one atom and this is truly a big deal and you might not think of at the moment but once this kind of Technology does become mainstream and we're able to make technology that is able to power AI That's as thin as a neuron and as effective as a brain you can really start to understand how quickly things are going to ramp up --- *Источник: https://ekstraktznaniy.ru/video/14873*