Showing papers by "Steven J. Plimpton published in 2018"

Multiscale Co-Design Analysis of Energy, Latency, Area, and Accuracy of a ReRAM Analog Neural Training Accelerator

Abstract: Neural networks are an increasingly attractive algorithm for natural language processing and pattern recognition. Deep networks with >50 M parameters are made possible by modern graphics processing unit clusters operating at $270\times $ energy and $540\times $ latency advantage over a similar block utilizing only digital ReRAM and takes only 11 fJ per multiply and accumulate. Compared with an SRAM-based accelerator, the energy is $430\times $ better and latency is $34\times $ better. Although training accuracy is degraded in the analog accelerator, several options to improve this are presented. The possible gains over a similar digital-only version of this accelerator block suggest that continued optimization of analog resistive memories is valuable. This detailed circuit and device analysis of a training accelerator may serve as a foundation for further architecture-level studies.

Effect of shape and friction on the packing and flow of granular materials

Abstract: The packing and flow of aspherical frictional particles are studied using discrete element simulations. Particles are superballs with shape ${|x|}^{s}+{|y|}^{s}+{|z|}^{s}=1$ that varies from sphere ($s=2$) to cube ($s=\ensuremath{\infty}$), constructed with an overlapping-sphere model. Both packing fraction, $\ensuremath{\phi}$, and coordination number, $z$, decrease monotonically with microscopic friction $\ensuremath{\mu}$, for all shapes. However, this decrease is more dramatic for larger $s$ due to a reduction in the fraction of face-face contacts with increasing friction. For flowing grains, the dynamic friction $\stackrel{\ifmmode \tilde{}\else \~{}\fi{}}{\ensuremath{\mu}}$---the ratio of shear to normal stresses---depends on shape, microscopic friction, and inertial number $I.$ For all shapes, $\stackrel{\ifmmode \tilde{}\else \~{}\fi{}}{\ensuremath{\mu}}$ grows from its quasistatic value ${\stackrel{\ifmmode \tilde{}\else \~{}\fi{}}{\ensuremath{\mu}}}_{0}$ as $(\stackrel{\ifmmode \tilde{}\else \~{}\fi{}}{\ensuremath{\mu}}\ensuremath{-}{\stackrel{\ifmmode \tilde{}\else \~{}\fi{}}{\ensuremath{\mu}}}_{0})=d{I}^{\ensuremath{\alpha}}$, with different universal behavior for frictional and frictionless shapes. For frictionless shapes the exponent $\ensuremath{\alpha}\ensuremath{\approx}0.5$ and prefactor $d\ensuremath{\approx}5{\stackrel{\ifmmode \tilde{}\else \~{}\fi{}}{\ensuremath{\mu}}}_{0}$ while for frictional shapes $\ensuremath{\alpha}\ensuremath{\approx}1$ and $d$ varies only slightly. The results highlight that the flow exponents are universal and are consistent for all the shapes simulated here.

Highly scalable discrete-particle simulations with novel coarse-graining: accessing the microscale

Abstract: Simulating energetic materials with complex microstructure is a grand challenge, where until recently, an inherent gap in computational capabilities had existed in modelling grain-scale effects at ...

Gas-kinetic simulation of sustained turbulence in minimal Couette flow

Abstract: Simulations of a gas in a Minimal Couette unit at Re=500 using Direct Simulation Monte Carlo, a molecular method enforcing molecular chaos for gas-molecule collisions, reproduces the turbulence from DNS. Thus molecular chaos does not prevent development of long-range correlations in turbulence.