Keeping Up With Modern Demands: Towards Power-efficient�  Embedded Systems

Ourania Spantidi

Ph.D. Candidate

Southern Illinois University, Carbondale, Illinois, USA

What’s the issue?

​

​

Which one do we prefer?

Achieving great performance? �Or saving some energy?

​

​

​

Let’s try to do both!

Approximate Computing

• Deep Neural Networks → can be computationally intense
• Embedded devices integrate hardware accelerators → MAC units

​

• Approximate Computing: trading accuracy for energy efficiency
• DNNs comprise many multiplications: target approximate multipliers

Reconfigurable Approximate Multiplier

5-step

Mapping Methodology

Deep Neural Network

Average energy gains

18.33%

Clustered Multiprocessors (CMPs)

• Utilize parameter mining and frequency scaling → energy gains!
• PSTL: What is the maximum power threshold we can have while maintaining thermal safety?

​

PSTL query

Parameter Mining

Applications

Average gains in power efficiency

11%

​

Run-time manager

Odroids XU-3 board