Dataset Information

Simulation of Inference Accuracy Using Realistic RRAM Devices.

ABSTRACT: Resistive Random Access Memory (RRAM) is a promising technology for power efficient hardware in applications of artificial intelligence (AI) and machine learning (ML) implemented in non-von Neumann architectures. However, there is an unanswered question if the device non-idealities preclude the use of RRAM devices in this potentially disruptive technology. Here we investigate the question for the case of inference. Using experimental results from silicon oxide (SiO x ) RRAM devices, that we use as proxies for physical weights, we demonstrate that acceptable accuracies in classification of handwritten digits (MNIST data set) can be achieved using non-ideal devices. We find that, for this test, the ratio of the high- and low-resistance device states is a crucial determinant of classification accuracy, with ~96.8% accuracy achievable for ratios >3, compared to ~97.3% accuracy achieved with ideal weights. Further, we investigate the effects of a finite number of discrete resistance states, sub-100% device yield, devices stuck at one of the resistance states, current/voltage non-linearities, programming non-linearities and device-to-device variability. Detailed analysis of the effects of the non-idealities will better inform the need for the optimization of particular device properties.

SUBMITTER: Mehonic A

PROVIDER: S-EPMC6582938 | biostudies-literature | 2019

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Simulation of Inference Accuracy Using Realistic RRAM Devices.

Mehonic Adnan A Joksas Dovydas D Ng Wing H WH Buckwell Mark M Kenyon Anthony J AJ

Frontiers in neuroscience 20190612

Resistive Random Access Memory (RRAM) is a promising technology for power efficient hardware in applications of artificial intelligence (AI) and machine learning (ML) implemented in non-von Neumann architectures. However, there is an unanswered question if the device non-idealities preclude the use of RRAM devices in this potentially disruptive technology. Here we investigate the question for the case of inference. Using experimental results from silicon oxide (SiO <sub><i>x</i></sub> ) RRAM dev ...[more]

PMID: 31249502

Similar Datasets

Project description:Abstract Gravity waves (GWs) generated by tropical convection are important for the simulation of large?scale atmospheric circulations, for example, the quasi?biennial oscillation (QBO), and small?scale phenomena like clear?air turbulence. However, the simulation of these waves still poses a challenge due to the inaccurate representation of convection, and the high computational costs of global, cloud?resolving models. Methods combining models with observations are needed to gain the necessary knowledge on GW generation, propagation, and dissipation so that we may encode this knowledge into fast parameterized physics for global weather and climate simulation or turbulence forecasting. We present a new method suitable for rapid simulation of realistic convective GWs. Here, we associate the profile of latent heating with two parameters: precipitation rate and cloud top height. Full?physics cloud?resolving WRF simulations are used to develop a lookup table for converting instantaneous radar precipitation rates and echo top measurements into a high?resolution, time?dependent latent heating field. The heating field from these simulations is then used to force an idealized dry version of the WRF model. We validate the method by comparing simulated precipitation rates and cloud tops with scanning radar observations and by comparing the GW field in the idealized simulations to satellite measurements. Our results suggest that including variable cloud top height in the derivation of the latent heating profiles leads to better representation of the GWs compared to using only the precipitation rate. The improvement is especially noticeable with respect to wave amplitudes. This improved representation also affects the forcing of GWs on large?scale circulation. Key Points Use of an idealized numerical model forced with radar?derived latent heat profiles to simulate convective gravity waves in the tropics Relating latent heat to rain rate and echo top height leads to improved representation of tropical GWs compared to satellite observations Approach could improve prediction of tropical gravity waves for weather forecasting and climate applications

Dataset Information

Simulation of Inference Accuracy Using Realistic RRAM Devices.

Publications

Simulation of Inference Accuracy Using Realistic RRAM Devices.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets