On June 12^th, 2023 at 3.30 pm Luca Terraciano, Research Engineer at the Huawei Research Center in Zurich, will give a seminar on "Huawei Ascend and the role of NPUs in modern AI and HPC applications" in DEIB Conference Room.

The amount of resources needed to enable modern AI tasks has grown significantly over the past few years. This growth has outperformed Moore's law and made modern computing systems incapable of supplying the rising resource demand. Thus, hardware accelerators started to gain more and more interests. They provide significant performance gains over traditional CPUs, making them an appealing option for data-intensive AI applications that require high computational power. Furthermore, offloading the compute-intensive tasks to specialized hardware enables the deployment of AI applications on a wider range of devices, including mobile and edge systems. This can enable new applications and use cases for AI, such as real-time object detection and recognition, natural language processing, and more.
This seminar offers an overview of the Huawei Ascend platform, the computing solution designed by Huawei for handling AI and tensor computation workloads. The Atlas devices target a wide range of execution environments and can be used in edge, cloud and HPC infrastructures. The talk will dive into the concept behind the Ascend devices, introduce the hardware architecture, and the programming models used to write operators and applications. To highlight the benefits of this technology, the talk will explore how Ascends can be used to speed up a particular workload: HPL-AI, a key benchmark for HPC systems with hardware accelerators. The seminar will discuss how to optimize performance from many perspectives: optimizing low-level operators, preventing PCI communication and HBM memory from becoming bottlenecks, and scaling out to multiple nodes.

Luca Terracciano received the Bachelor of Science (2018) and the Master of Science degree (2021) in Computer Engineering from Politecnico di Milano. He is currently working at the Huawei Research Center in Zurich as a Research Engineer, where he primarily focuses on heterogeneous scheduling techniques and task-based parallelism in High Performance Computing applications. His other areas of interest are self-adaptive systems and autoscaling techniques in cloud and edge computing applications.