Power-Efficient AI at the Edge: Real-World Gains with Model Sparsity
(Room 303B)
04 Nov 25
3:25 PM
-
3:50 PM
Tracks:
Embedded AI: AI Efficiency
As edge AI applications demand ever-higher performance with tight power constraints, neural network optimization is becoming essential. One powerful technique is unstructured sparsity, which involves removing redundant or less significant weights from a trained model. This results in models that require significantly less processing power, with minimal to no loss in accuracy, and can lead to substantial power consumption savings.
Compared to other model compression techniques like quantization, which reduces the precision of weights and activations, sparsity reduces the number of active computations altogether. When effectively combined, these techniques can deliver up to 10X performance gains and up to 90% power savings, in addition to significantly reducing memory requirements.
This session will explore the real-life benefits of unstructured sparsity, highlighting performance and power consumption improvements achieved with state-of-the-art networks such as YOLOv8/9. We'll also examine its applicability to edge AI processing and machine vision use cases. Attendees will gain insight into how pruning techniques can be applied to real-world ML deployments using existing framework capabilities, unlocking new potential across industrial, automotive, and IoT applications.