Home
Lectures
Platform for Next-Generation Analog AI Hardware Acceleration

Platform for Next-Generation Analog AI Hardware Acceleration

Speaker: Kaoutar El Maghraoui – Yorktown Heights, NY, United States
Topic(s): Artificial Intelligence, Machine Learning, Computer Vision, Natural language processing , Hardware, Power and Energy

Abstract

Analog In-Memory Computing (AIMC) is a game-changing approach that boosts the efficiency of Deep Neural Network (DNN) inference and training. It tackles the performance losses caused by data movement between computational units and memory. AIMC harnesses memristive crossbar arrays to reduce the Von Neumann bottleneck, enabling highly parallel computations directly in memory. However, AIMC isn't without its challenges. The inherent noise, non-linear device characteristics, and non-ideal peripheral circuitry demand sophisticated algorithmic adaptations to maintain accuracy comparable to digital systems.

To address these challenges, IBM Research has introduced the IBM Analog Hardware Acceleration Kit (AIHWKit), a pioneering open-source toolkit integrated within PyTorch (available at https://github.com/IBM/aihwkit). This toolkit is designed to simulate crossbar arrays, allowing users to assess the impact of material properties and non-idealities on DNN accuracy. It supports advanced functionalities, including hardware-aware training, mixed-precision training, and advanced analog training optimizers. The AIHWKit also facilitates inference and training on real research Phase-change memory (PCM)-based analog AI chip prototypes.

In this lecture, we will explore the concepts of IMC through the capabilities and applications of the AIHWKit and its cloud front end. We will show how this platform can model and optimize DNNs for deployment on AIMC platforms. Attendees will learn how to utilize the web front-end cloud composer of the AIHWKit, which offers a no-code experience and interactive demos to configure experiments, simulate training and inference, and access inference IMC hardware prototypes. This lecture aims to equip participants with the skills to understand, model, simulate, and enhance the training and inference of complex DNNs using AIMC, paving the way for more energy-efficient and faster AI solutions.

About this Lecture

Number of Slides: 65
Duration: 60 minutes
Languages Available: English
Last Updated:

Request this Lecture

To request this particular lecture, please complete this online form.

Request a Tour

To request a tour with this speaker, please complete this online form.

All requests will be sent to ACM headquarters for review.