N-ary crossbar architecture enables multibit neural inference
A simulation framework designed for N-ary crossbar architectures enables neural network inference with few assumptions. The tasks of XOR and MNIST classification were executed using a simulated 4-state magnetic tunnel junction (MTJ) crossbar array configured as 4x4. The accuracy for MNIST reached 94.48%, which is lower than the 97.56% achieved by software. To reduce the performance gap between software and hardware, PCA dimensionality reduction was utilized. It was determined that weight quantization was the main source of error, which was analyzed alongside systematic nonidealities and random noise. Interestingly, cell-specific random noise was found to be less harmful than systematic errors due to the averaging effect across the array. An optimal balance of states per cell to minimize quantization error was also demonstrated.
Key facts
- Simulation framework for N-ary crossbar architectures
- XOR and MNIST classification tasks inferred
- Simulated (4x4) 4-state MTJ crossbar array used
- MNIST accuracy: 94.48% vs 97.56% software baseline
- PCA dimensionality reduction reduced performance gap
- Weight quantization identified as primary error source
- Cell-specific random noise less detrimental than systematic errors
- Optimal number of states per cell demonstrated
Entities
Institutions
- arXiv