Overview

The AIHWKIT addresses the growing demand for an alternative to conventional digital computing hardware, such as CPUs and GPUs, which are energy-intensive and less efficient for AI workloads. For example, the automotive, healthcare, telecommunications, and consumer electronics sectors increasingly require energy-efficient alternatives to conventional computing for applications such as video processing, natural language processing, and machine learning workloads, both in cloud environments and at the edge.

Our result

The Analog Hardware Acceleration Kit (AIHWKIT) is an advanced tool developed to accelerate artificial intelligence (AI) research by facilitating the simulation and training of neural networks on analog hardware. HYBRAIN leverages the AIHWKIT to train neural networks that are subsequently deployed on custom analog hardware, including Analog In-memory Compute (AIMC) tiles. The noise modelling capabilities of the AIHWKIT help ensure a smooth transition from simulated models to physical deployment, allowing for precise evaluation of system performance and robustness. Under the HYBRAIN project, further enhancements have been made, including the creation of a purely PyTorch-based version of the computational “tile.” This PyTorch-based version is particularly important as it seamlessly integrates with the widely used PyTorch framework, simplifying the use of existing PyTorch tools and libraries for researchers and developers. It also enhances compatibility, reduces development time, and provides a more intuitive and accessible implementation compared to prior versions.

Needs and target audiences

The AIHWKIT is a powerful tool for hardware-aware training of neural networks implemented on analog computing systems. One of its unique features is its ability to benchmark analog computing hardware against traditional digital equivalents across a variety of neural network models, offering an accurate evaluation of hardware performance and trade-offs. We specifically need researchers and engineers working on analog AI hardware, noise-aware neural network training, and hardware/software co-design for efficient AI systems.

Current stage

The central market challenge addressed by the AIHWKIT is training neural networks that remain robust under the influence of noise introduced by analog in-memory computing hardware. Analog hardware offers substantial energy efficiency and scalability but is limited by its inherent noise, which affects computational precision, especially during inference.

Unique value proposition

The AIHWKIT is a powerful tool for hardware-aware training of neural networks implemented on analog computing systems. One of its unique features is its ability to benchmark analog computing hardware against traditional digital equivalents across a variety of neural network models, offering an accurate evaluation of hardware performance and trade-offs.

Keywords

• brain-inspired computing
• physical-substrate computing
• in-memory computing
• unconventional computing
• neuromorphic computing

Research project acronyms

• HYBRAIN

Research projects

• Hybrid electronic-photonic architectures for brain-inspired computing

Project identifiers

• 101046878

Contributors

• UNIVERSITEIT TWENTE

Source: Horizon Results Platform