Benchmarking Neuromorphic Hardware and Its Energy Expenditure

Area of Science:

• Neuromorphic Engineering
• Computer Architecture
• Artificial Intelligence Hardware

Background:

• Neuromorphic hardware aims to mimic the brain's efficiency but lacks standardized evaluation methods.
• Comparing diverse neuromorphic architectures (mixed-signal, digital) requires a comprehensive benchmarking approach.

Purpose of the Study:

• To propose and discuss a platform-overarching benchmark suite for neuromorphic hardware.
• To enable standardized comparison of different neuromorphic systems.
• To assess the energy efficiency gap between current neuromorphic hardware and the human brain.

Main Methods:

• Development of a benchmark suite covering low-level characterization to high-level application evaluation.
• Utilizing benchmark-specific metrics for performance and energy efficiency.
• Introduction of a predictive energy model for neuromorphic systems.

Main Results:

• Characteristic performance differences were revealed across various neuromorphic hardware platforms.
• Current neuromorphic systems are estimated to be at least four orders of magnitude less energy-efficient than the human brain.
• A significant efficiency gap (two to three orders of magnitude) remains even with modern fabrication processes.

Conclusions:

• The proposed benchmark suite provides a standardized method for evaluating neuromorphic hardware.
• There is a substantial need for further advancements in neuromorphic computing to approach biological brain efficiency.
• Comparison with standard computing approaches highlights the potential and challenges of neuromorphic solutions.