Content area
With the evolving AI landscape there is a growing need for larger models with several trillions of parameters, challenging storage and compute intensity solutions. There are narrowing innovation opportunities in Dynamic Random Access Memory (DRAM) for increasing storage densities through advanced packaging. This includes enhanced architectures for bandwidth increase - with even more challenge at the edge. Hence, the significant research growth in analog-based neural networks to overcome these growing challenges. However, the promise of decreased latencies, increased computational parallelism, and higher storage densities is not without caveats.
This dissertation evaluates current and potential methods for reliably programming Hexagonal Boron-Nitride (h-BN) memristors from the perspective of crossbar arrays by emulating proximity to drivers through varying amplitude and slew rate. It goes beyond previous work to include experimental understanding of the dynamic conductive nano-filament formation upon programming. A comprehensive study of the temporal response under various conditions – voltage pulse amplitude, edge rate (pulse rise/fall times), and temperature – provides new insights into the resistive switching process towards optimized devices and improvements in their implementation in analog based neural networks. The results show that h-BN memristors can achieve multi-state operation through ultrafast pulsed switching (< 25 ns), and high energy efficiency (~ 10 pJ/pulse).
Addressing program induced variability (PIV) and noise (random telegraph noise or RTN) is a persistent concern for conventional resistive random-access memory (RRAM), especially for multi-state operation. The effects on accuracies of a neural network classifier were evaluated and it was seen that a larger contribution of errors came from PIV (5 - 10% more) over RTN. Experimental results on h-BN memristors reveal a lowering in RTN fluctuation when programmed to higher conductance states (80% to 4% reduction in RTN fluctuation measured on h-BN memristors as conductance varied from 45 nS to 250 µS ranges). Nonetheless, the impact of RTN can be exacerbated in memristors prepared using common transfer methods (i.e., h-BN film transferred from a separate growth substrate onto the processing substrate). A comparison with a novel approach (direct growth on a substrate) is conducted in this work showing less RTN fluctuation and resulting in up to 70% less variation in neural network classifier accuracies.