The Strong Sensitivity of the Characteristics of Binary Stochastic Neurons Employing Low Barrier Nanomagnets to Small Geometrical Variations

Binary stochastic neurons (BSNs) are excellent activators for machine learning. An ideal platform implementing them low- or zero-energy-barrier nanomagnets (LBMs) possessing in-plane anisotropy (e.g. circular slightly elliptical disks) whose fluctuating magnetization encodes a probabilistic (p-) bit. Here, we show that such BSN's activation function, the pinning current (which pins output to particular binary state), and response time - all exhibit strong sensitivity very slight geometric variations in LBM's cross section. A mere 1% change diameter of nanomagnet any arbitrary direction can alter by factor ~4 at room temperature 10% ~2. All this causes large device-to-device variation which is detrimental integration. We also energy dissipation lowered but increased replacing cross-section with one then encoding p-bit component along major axis. Encoding minor axis has opposite effect. The energy-delay-product, however, relatively independent whether circle an ellipse case ellipse.