AI on PC: Local, Private, Instant Intelligence by Vinod Kathail

Abstract
AI/ML workloads are increasingly shifting from cloud-only deployment to a hybrid model in which the PC becomes an intelligent endpoint executing AI models locally. This talk focuses on the local execution of AI models and applications on PCs, discussing why on-device execution matters, key client-side use cases, and how on-device models can interoperate with cloud-hosted models to deliver end-to-end experiences. Modern AI PCs integrate dedicated accelerators, commonly called Neural Processing Units (NPUs), alongside the CPU and GPU. We will use the NPU in AMD’s Ryzen AI devices as an example to illustrate their power-efficient architecture and then walk through the end-to-end software stack for seamless model deployment.

Biography
Vinod Kathail is a Senior Fellow at AMD and a Visiting Professor at IITD. His current focus is on the end-to-end AI/ML software stack for NPUs in Ryzen AI and embedded devices, building on his work at Xilinx. At Xilinx, he was a Fellow and Chief Architect for the Vitis high-level programming environment. Earlier, he co-founded Synfora, a high-level synthesis company, and served as its CTO. Prior to Synfora, he worked at HP Labs, where he led the PICO high-level synthesis project. He was also one of the architects of the VLIW architecture and compiler that became Intel Itanium. He has numerous patents and publications in the areas of  AI accelerators and their SW stack, parallel/heterogeneous architectures, parallel programming environments, and high-level synthesis. He received a Doctor of Science in EECS from MIT.