Relative Efficiency of Memristive and Digital Neuromorphic Crossbars
| Autor: | Christopher D. Krieger, Mark McLean, David J. Mountain |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
010302 applied physics
Artificial neural network Computer science Computation 020208 electrical & electronic engineering Spice 02 engineering and technology Memristor 01 natural sciences law.invention Efficiency Neuromorphic engineering Computer architecture law 0103 physical sciences 0202 electrical engineering electronic engineering information engineering Throughput (business) Efficient energy use |
| Zdroj: | Proceedings of the International Conference on Neuromorphic Systems. |
| DOI: | 10.1145/3229884.3229887 |
| Popis: | Machine learning is becoming an increasingly large consumer of compute resources. This demand has led to interest in neuromorphic processors (NMPs) that can quickly and efficiently perform learning and inference tasks. Both analog and digital NMPs have been proposed in literature, with many papers indicating that memristor based analog designs provide significant efficiency gains over digital approaches. In this work, we use extensive SPICE simulations to compare the energy efficiency and throughput per area of basic analog and digital neuromorphic processors after architectural differences have been factored out. We find that for our test system, memristor-based analog designs offer about four times better throughput per watt for computation than digital systems and that they provide nearly ten times the throughput per area. If these designs were used in a complete system, including memory and I/O, the gains from using memristors would be small. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|