 From around the globe, it's theCUBE with digital coverage of AWS re-invent 2020, sponsored by Intel, AWS, and our community partners. Welcome back to theCUBE's virtual coverage of AWS re-invent 2020. It's virtual this year, we're not in person, so we're doing remote interviews. Part of the three weeks we'll be covering wall to wall, a lot of great conversations, news to cover, and joining me today, fresh off the news of Andy Jassy's keynote. We have two great guests here, Chaitan Kapoor, Senior Product Manager for Accelerated Computing at AWS, and A-Time Medina, Chief Business Officer at Habana Labs, which was recently acquired by Intel. Folks, thanks for coming on, gentlemen. Thank you for spending the time, for coming on theCUBE, appreciate it. Thanks for having us. Chaitan, so talk about the news. Obviously, compute is changing, it's being re-invented. That's the theme from Andy's keynote. What did Andy announce? Could you take a minute to explain the announcement? What services, what API, what's going to be supported, what's this about, take a minute to explain. Yeah, absolutely. Yeah, so today we announced our plans to launch an EC2 instance based on hardware accelerators from Habana Labs. We expect these instances to be available in the first half of next year. And these are custom designed for accelerating training of deep learning models. As we all know, training of deep learning models is a really computationally extensive task, oftentimes it takes too long and costs too much. And we're really excited about getting these instances out of the market, as we expect for them to provide up to 40% better price performance than top of the line GPU instances. A lot of improvements. Why did AWS do this? Why Habana? What's the working backwards document tell you? What is it customers looking for here? Is there a specific use case? Yeah, absolutely. Over the years, the use of machine learning and deep learning has really skyrocketed. So we are seeing companies from all the way like Fortune 500 to startups just reinventing their business models and using deep learning more pervasively. So we have companies like Pinterest, who use deep learning for content recommendations and object detection to Toyota research institute that are advancing the science behind autonomous vehicles. And there's a consistent team from a lot of these customers that are innovating in the deep learning space that the cost it takes to experiment, train and optimize the deep learning models is too high. And they're looking at us as one of their partners to help them optimize their costs, bring them as low as possible while giving them really performing products and enable them to actually bring their markets, their innovations to market as soon as possible, right? So to answer your questions straight on, you know, what's working backwards? It's a feedback from customers that they want choice and they want our help to lower the amount of compute resources and the cost it takes to train the deep learning models. Hey, Tom, why don't you weigh in here on Hibana and now part of Intel. What trends are driving this? What's the motivation? Where you guys fit in? What's the view on this? Yeah, so Hibana was founded in 2016 to deliver AI processors for the data center and cloud for training and inference deep learning models. So while building chips is hard, building the software and ecosystem is even harder. So joining forces with Intel simply helps us connect the dots. Ever since the acquisition last year, we were able to significantly boost our R&D resources. And now we're leveraging Intel scale in number of customers and ecosystem and partner support. So what's the name of the product? Is there a chip name? Was it Gaudi is the name? Yes, the product is Hibana Gaudi. Okay, and so it's going to be hardware. So it's a hardware software. What's involved? Take us through the product. Yes, so Gaudi was designed from the ground up to do one task, which is training deep learning models. To do that well, we focused the architecture on two aspects, efficiency and scalability. The compute architecture is a combination of fully programmable TPC tensor processor cores and a central GM engine. These TPC cores are programmable VLIW SIMD machines that we designed with custom instruction set architecture and special functions that will develop specifically for ADI. The Gaudi chip integrates also 32 gigabyte of HBM2 memory, which makes it easy to port to for GPU developers. Gaudi is unique in integrating 10 ports of 100 gigabit internet rocky on chip. And this is opposed to other architecture which use proprietary interfaces. So overall improving the cost performance is achieved through efficiency, namely higher utilization of the compute and memory resources on chip and the native integration of the rocky interfaces. JTown, this is actually interesting. This is the theme for reinvent, we're seeing it right on stage today, play out again, another command performance by Andy Jassy, slew of announcements. How does Gaudi fit into the AI portfolio or Amazon strategy? Because what Aitown's saying is, it sounds like he's doing the heavy lifting on all this training stuff when people want to just get to the outcome. I mean, the theme has been just let the products do what they do, kind of put stuff under the covers and just let it scale. Is that the theme here? What does this all fit in? Take us through how this fits into the AI strategy for Amazon and also what does Hibana Intel bring to the table? Absolutely. Yeah, so with respect to our overall strategy and portfolio units, it's relatively straightforward, right? So we are laser focused on making sure we have the broadest and the deepest portfolio of services for machine learning, right? So these range from infrastructure services, specifically compute, networking and storage, all the way up to like managed ML services which come with pre-trained models and customers can simply invoke them using an API call, right? So from a strategy perspective, we want to make sure that we provide a customer to a choice, enable them to pick the right platform for the right use case, help them get to the cost structure they actually want, right? So with Hibana and their acquisition with Intel, we finally have access to hardware, software and the ability to kind of build out an ecosystem beyond what, you know, traditionally has been used, which is GPUs, right? So the engagement with Hibana, you know, allows us to take their products and capabilities, wrap it around an EC2 instance, which is what customers will be able to launch, right? And doing so, we're enabling them to, you know, tap into the innovation that Eton and the rest of the Hibana team are working on while having a solution that is integrated with the full AWS stack, right? So you don't have to rack and stack hardware in your data center. These are going to be available as standard EC2 instances. You can just click and launch them, get access to software that's already pre-integrated and baked in and ready to go, right? So it actually comes down to taking their innovations, coupling it with an AWS solution and making it super easy for customers to gather up and running with respect to training the deep web models. All right, well, here's the question that I want to get to, I think everyone's, on everyone's mind is, how is it gaudy different or similar than other GPUs specifically? You mentioned the software stack on AWS. Would you get the software stack inside the chip? How is this different or similar to other GPUs and what's the difference between the software stack versus say traditional libraries? So from day one, we were focused on the software experience and we were mindful in the need to make it easy for developers to use the innovations we have in the hardware. Most developers, if not all of them, are using deep learning frameworks such as TensorFlow and PyTorch for building their deep learning models. So Gaudy Synapse AI software suite comes integrated and optimized for TensorFlow and PyTorch. So we expect most developers to be able to take their existing models and with minor changes to the training streets to be able to run them on Gaudy based instances. In addition, expert developers that are familiar with writing their own kernels will be provided with a full tool suite for writing their own TPC kernels that can augment the Habana provided library. So that's a user experience for the developers, right? That's what you're saying? Exactly, exactly. And we will provide detailed guides for developers in doing that. Habana will provide open access to documentation, library, software, models, and other to Habana's GitHub and bidirectional communication with the Habana developer community. All these resources will be available concurrently with the AWS Instances launch. Okay, so if I'm a developer, how do I get involved? It's software on GitHub. I use the hardware on Amazon, obviously in their instances, it's a new instance. Take me through the workflow. I'm a developer. I'm into this. I want to get involved. What am I doing? Take me through. Yeah, so I think it's, so if the developer is accustomed to using GPUs for training, the deep learning models, the experience is going to be practically the same, right? So they'll have multiple options to get started. One of them would be, for example, to take our deep learning AMIs or Amazon machine images that'll come integrated with software from Habana Labs, right? So customers will take the deep learning AMI and launch it on an EC2 instance, featuring the Gaudi accelerators, right? So when, with that, they'll have the baseline construct of software and hardware available to get up and running with, right? We'll support all different types of workflows. So if customers want to use containerized solutions, these instances will be supported via our ECS and EKS services. So if they're using containerized Kubernetes, these, the solution will just work. And lastly, we also intend to support these instances through SageMaker. So just a quick recap on SageMaker, that's a managed service that does end-to-end, that provides end-to-end capabilities for training, debugging, building and deploying machine learning applications. So these instances will also be supported in SageMaker. So if you're familiar with SageMaker, you can get up and running with the instances fairly quickly. Sounds like it's going to enable a lot of action at SageMaker level and then can that layer. On the use cases, I got to ask you guys quickly, what's the low-hanging fruit use case applications for this product, this partnership? Because that's going to be the first traction set. What are some of these applications going to be useful? What can we expect to see? So typical applications would be image classification, object detection, natural language processing, the recommendation systems. You'll find reference models in our GitHub for that and we'll be growing that list as you can imagine. Okay, where can people find more info? Give us the data, take a minute to explain, put a plug in for how, what's all the coordinates, URL, site support, how do people create, how do people get involved in the community? Yeah, so customers are going to be able to access information on AWS websites and also on Habana Labs websites. So we'll be kicking off a preview early next year. So I would highly recommend for customers to find our product pages and sign up for early access and preview information. You, Tom? Yes, and you'll find more information on Habana.ai as well as Habana's GitHub over there. Great announcement, congratulations. Thanks for sharing the news and some commentary on it. This is really the big theme, you know, at COVID-19 and this pandemic has shown as massive acceleration of digital transformation and having the software and hardware out there that accelerates the heavy lifting and creates value around the data, super valuable. Thanks for doing that, appreciate taking the time. Thank you so much. Yeah, thanks for having us. Okay, this is theCUBE's coverage at AWS re-invent. Next three weeks we're here on the ground, we're remote, we're live, we're inside the studio. We wish we could be there in person, but it's remote this year, but stay tuned. Check out siliconangle.com, exclusive interviews with Andy Jassy and Amazon executives and the big news covering there all there in one spot, check it out. We'll be back with more coverage after this break. Thanks for watching.