CESG Eminent Scholar Series:
Balint Fleischer of Huawei’s Central Research Institute
“CPU and Server System Architecture Opportunities for AI Application Optimization”
Abstract:For the past 50 years, the computer industry has been focusing on improving transactional workloads. We are now seeing the emergence of a new class of “Narrow AI” based on applications playing an increasingly critical role in diverse use cases from Robotics, Smart Cities, Expert Systems, Medical Diagnostics, Financial Systems to Research and so forth. They perform assistive functions through speech recognition, face and image recognition, Fraud Detection, retrieving complex data structures and the integration of diverse information. AI applications are fundamentally different from classic applications. Classic applications are based on explicit programming using arithmetic and logic operations, while AI applications are trainable or self-learning algorithms to make predictions. AI applications use heterogeneous streaming data as opposed to classic applications, which transactional and structured data. Classic CPU architectures are very inefficient for AI applications; they lack sufficient memory BW for a diverse set of accelerators to emerge. However, E2E application “pipelines” are a hybrid requiring the creation of a new server platform capable of efficiently supporting new use cases. This presentation will highlight some of the ongoing development in this area and what could be the future direction.
Bio: Balint Fleischer is currently Chief Scientist at Huawei’s Central Research Institute, where he is responsible for research into next generation data center and server architectures. He was most recently CTO at startup Parallel Machines, where he developed new architectures for advancing predictive analytics and machine learning. Previously he was the General Manager and Director of Architecture development, including efforts related to 3DXPoint and Rack Scale Architecture. He also had a long residency at Sun Microsystems including being VP/CTO of the Networked Storage Division, where he led the design of next generation storage systems and storage virtualization platforms; while at Sun he led Sun’s architecture development for many successful low end midrange server products and was responsible for the company’s InfiniBrand effort focusing on enterprise clustering, I/O, and storage.