There has been tremendous demand to deploy AI models across new and diverse hardware architectures. Many of these architectures include a variety of processing nodes and specialized hardware accelerators. The challenge is to take trained AI models developed in various open-source software frameworks and execute them efficiently on these architecures. Software tools must evolve to provide features such as AI model import, graph analysis, quantization, optimization, and code generation. Creating AI-centric software development tools is a complex undertaking that requires expertise in AI network theory and construction, high-performance computing, compilers, and embedded systems. This talk will share some of our experiences developing Cadence’s NeuroWeave Software Development Kit (SDK). NeuroWeave is a collection of tools and software libraries for optimizing and compiling AI models in order to execute them efficiently on Tensilica DSPs and Neo accelerators.
Eric Stotzer
Eric Stotzer is a Software Engineering Group Director at Cadence Design Systems, where he is responsible for the Neuroweave SDK and Xtensa Neural Network Compiler (XNNC). Eric worked for 30 years at Texas Instruments developing system software tools for DSPs and MCUs. Before coming to Cadence, he was at Mythic working on a neural network compiler for mixed-signal AI accelerators. He is a coauthor of the book Using OpenMP - The Next Step: Affinity, Accelerators, Tasking, and SIMD, (MIT Press 2017). Eric holds a PhD in Computer Science from the University of Houston.