Fair Energy Comparison Between SNNs and QNNs
A new paper challenges simplified energy evaluations of Spiking Neural Networks (SNNs) by introducing a rigorous analytical model that accounts for data movements and memory accesses. The authors map rate-encoded SNNs with T timesteps to capacity-matched Quantized Artificial Neural Networks (QNNs) with ⌈log₂(T+1)⌉ bits, ensuring comparable representational capacities and hardware requirements. Their detailed energy model covers core computation and data movement, providing a fair baseline for comparing the true energy benefits of SNNs over QNNs. The work highlights that neglecting overheads can lead to misleading conclusions about SNN efficiency.
Key facts
- SNNs promise higher energy efficiency than QNNs due to event-driven computation.
- Prevailing energy evaluations oversimplify by focusing on computation and neglecting data movements and memory accesses.
- The paper maps rate-encoded SNNs with T timesteps to QNNs with ⌈log₂(T+1)⌉ bits.
- This mapping ensures comparable representational capacities and hardware requirements.
- A detailed analytical energy model encompassing core computation and data movements is introduced.
- The model enables meaningful energy comparisons between SNNs and QNNs.
- The paper is available on arXiv with ID 2409.08290.
- The announcement type is replace-cross.
Entities
Institutions
- arXiv