[XDF 2019] KeyNote 10 - Salil Raje EVP

Welcome to the Xilinx Adapt event. We have an amazing diversity of users and applications, from robotic surgery to communication equipment to data analytics. Our goal is to offer you something new and exciting so you can build even cooler applications. We have a comprehensive agenda with 76 hours of sessions, labs, and 69 members. I'm excited to talk about Vitis, our unified software development platform, a multi-year project that has been a personal passion of mine. It's a software product for software developers, data scientists, and AI engineers, giving you the best out of our adaptable devices. We will continually improve to keep up with increasing design complexity. Today we focus on Vitis and show how we welcome newcomers. I'll give context on what guided Vitis development: industry trends toward heterogeneous computing, applications spanning from cloud to edge, and the proliferation of AI. Standard CPUs cannot provide the performance scaling needed for compute-intensive workloads, so domain-specific accelerators like FPGAs and ACAPs are gaining momentum. The challenge is programming and integrating these heterogeneous accelerators, especially for software developers. Another trend is applications moving from cloud to edge for latency and privacy. For example, a surveillance system can start with a dumb camera and cloud processing, then move object detection to the edge, and finally make the camera smart. Vitis addresses the challenge of retargeting applications across cloud and edge. AI is being used in smart cities, autonomous driving, and more. Data scientists need to implement machine learning algorithms from frameworks like TensorFlow and PyTorch onto adaptable devices and integrate them into the full application. These three trends are intertwined, requiring a unified development environment. Vitis integrates all our software tools—SDx, SDAccel, ML Suite—into one platform alongside Vivado. Vitis enables many developers and data scientists to deploy applications on Zynq-based systems, ACAP-based systems, or in the cloud. Vitis is a common solution to build, run, and analyze applications. Development is done in the context of a platform, a pre-configured system with memory, interfaces, and software APIs. The host CPU communicates to the target device. Vitis offers dedicated compilers for C/C++ targeting programmable logic and the processing subsystem, plus system compilers and linkers. It integrates an x86 toolchain for PCIe systems and an AI engine compiler for Versal. The runtime, XRT, is OS-independent and platform-independent, and it's open source on GitHub. You can analyze your application with the Vitis performance analyzer to improve partitioning and performance. We also provide pre-optimized libraries—over 400 functions, some domain-specific like video and data analytics, others cross-domain like graph algorithms and DSP. All libraries are open source. You can target Zynq platforms using XRT, repackage for ACAP platforms, or scale up with Docker and Kubernetes. For AI, Vitis AI provides a way to deploy inference engines using DPUs (deep learning processing units), which are domain-specific accelerators for machine learning. You can download DPUs for Versal or Xilinx devices, compile your TensorFlow or PyTorch models through the Vitis AI stack, and achieve 10-20x performance improvement. Vitis AI is fully integrated with Vitis for whole-application acceleration. Let me bring it all together with an example: a smart city surveillance system. In a cloud-based system, you have a Versal accelerator card connected to a Zynq, using Vitis AI for machine learning and Vitis for computer vision. You can map the same application to a single ACAP device, moving the video analytics to the processing subsystem and the camera feed to the ACAP, creating an embedded device. In another example, a CPU+GPU configuration handling 7 channels of 1080p video can be moved to an ACAP solution handling 32 channels—a 4.5x improvement in throughput at one-sixth the power. Whole-application acceleration is extremely important for ACAP devices and Vitis. We have invested heavily in tools, runtime, and libraries, but the solution is not complete without comprehensive training and documentation. We have launched a developer website, developer.xilinx.com, with 20 articles and 30 projects, which will be the single hub for Vitis developers. Now let's invite on stage some of our customers to talk about how Xilinx technology and adaptable acceleration are helping them. Please welcome the VP of AI at IBM Cognitive Systems.