-
Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators
Authors:
Malte J. Rasch,
Charles Mackin,
Manuel Le Gallo,
An Chen,
Andrea Fasoli,
Frederic Odermatt,
Ning Li,
S. R. Nandakumar,
Pritish Narayanan,
Hsinyu Tsai,
Geoffrey W. Burr,
Abu Sebastian,
Vijay Narayanan
Abstract:
Analog in-memory computing (AIMC) -- a promising approach for energy-efficient acceleration of deep learning workloads -- computes matrix-vector multiplications (MVMs) but only approximately, due to nonidealities that often are non-deterministic or nonlinear. This can adversely impact the achievable deep neural network (DNN) inference accuracy as compared to a conventional floating point (FP) impl…
▽ More
Analog in-memory computing (AIMC) -- a promising approach for energy-efficient acceleration of deep learning workloads -- computes matrix-vector multiplications (MVMs) but only approximately, due to nonidealities that often are non-deterministic or nonlinear. This can adversely impact the achievable deep neural network (DNN) inference accuracy as compared to a conventional floating point (FP) implementation. While retraining has previously been suggested to improve robustness, prior work has explored only a few DNN topologies, using disparate and overly simplified AIMC hardware models. Here, we use hardware-aware (HWA) training to systematically examine the accuracy of AIMC for multiple common artificial intelligence (AI) workloads across multiple DNN topologies, and investigate sensitivity and robustness to a broad set of nonidealities. By introducing a new and highly realistic AIMC crossbar-model, we improve significantly on earlier retraining approaches. We show that many large-scale DNNs of various topologies, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers, can in fact be successfully retrained to show iso-accuracy on AIMC. Our results further suggest that AIMC nonidealities that add noise to the inputs or outputs, not the weights, have the largest impact on DNN accuracy, and that RNNs are particularly robust to all nonidealities.
△ Less
Submitted 16 February, 2023;
originally announced February 2023.
-
Evolutionary optimization of all-dielectric magnetic nanoantennas
Authors:
Nicolas Bonod,
Sebastien Bidault,
Geoffrey W. Burr,
Mathieu Mivelle
Abstract:
Magnetic light and matter interactions are generally too weak to be detected, studied and applied technologically. However, if one can increase the magnetic power density of light by several orders of magnitude, the coupling between magnetic light and matter could become of the same order of magnitude as the coupling with its electric counterpart. For that purpose, photonic nanoantennas have been…
▽ More
Magnetic light and matter interactions are generally too weak to be detected, studied and applied technologically. However, if one can increase the magnetic power density of light by several orders of magnitude, the coupling between magnetic light and matter could become of the same order of magnitude as the coupling with its electric counterpart. For that purpose, photonic nanoantennas have been proposed, and in particular dielectric nanostructures, to engineer strong local magnetic field and therefore increase the probability of magnetic interactions. Unfortunately, dielectric designs suffer from physical limitations that confine the magnetic hot spot in the core of the material itself, preventing experimental and technological implementations. Here, we demonstrate that evolutionary algorithms can overcome such limitations by designing new dielectric photonic nanoantennas, able to increase and extract the optical magnetic field from high refractive index materials. We also demonstrate that the magnetic power density in an evolutionary optimized dielectric nanostructure can be increased by a factor 5 compared to state of the art dielectric nanoantennas. In addition, we show that the fine details of the nanostructure are not critical in reaching these aforementioned features, as long as the general shape of the motif is maintained. This advocates for the feasibility of nanofabricating the optimized antennas experimentally and their subsequent application. By designing all dielectric magnetic antennas that feature local magnetic hot-spots outside of high refractive index materials, this work highlights the potential of evolutionary methods to fill the gap between electric and magnetic light-matter interactions, opening up new possibilities in many research fields.
△ Less
Submitted 21 January, 2019;
originally announced January 2019.
-
Fiber-integrated single photon source of high efficiency based on a concept of ultra-broadband optical horn antenna
Authors:
T. Grosjean,
M. Mivelle,
G. W. Burr,
F. I. Baida
Abstract:
We theoretically demonstrate a fiber-integrated single photon source of unprecedented efficiency. This fiber single photon source is achieved by coupling optically a single quantum emitter to a monomode optical fiber with a new concept of ultra-broadband optical antenna. Such an optical antenna concept is the result of the transposition to optical frequencies of the well-known low-frequency horn a…
▽ More
We theoretically demonstrate a fiber-integrated single photon source of unprecedented efficiency. This fiber single photon source is achieved by coupling optically a single quantum emitter to a monomode optical fiber with a new concept of ultra-broadband optical antenna. Such an optical antenna concept is the result of the transposition to optical frequencies of the well-known low-frequency horn antenna The optical horn antenna is here shown to be capable of directing the radiation from the emitter toward the optical fiber and efficiently phase-matching the photon emission with the fiber mode. Numerical results show that an optical horn antenna can funnel up to 85% of the radiation from a dipolar source within an emission cone semi-angle as small as 7 degrees (antenna directivity of 300). It is also shown that 50% of the emitted power from the dipolar source can be collected and coupled to an SMF-28 fiber mode over spectral ranges larger than 1000 nm, with a maximum energy transfer reaching 70 %. This approach is highly promising in the engineering of all-fiber "on-demand" single photon sources for telecommunication, quantum optics and sensing.
△ Less
Submitted 31 March, 2016; v1 submitted 6 March, 2016;
originally announced March 2016.
-
Phase change memory technology
Authors:
Geoffrey W. Burr,
Matthew J. Breitwisch,
Michele Franceschini,
Davide Garetto,
Kailash Gopalakrishnan,
Bryan Jackson,
Bulent Kurdi,
Chung Lam,
Luis A. Lastras,
Alvaro Padilla,
Bipin Rajendran,
Simone Raoux,
Rohit S. Shenoy
Abstract:
We survey the current state of phase change memory (PCM), a non-volatile solid-state memory technology built around the large electrical contrast between the highly-resistive amorphous and highly-conductive crystalline states in so-called phase change materials. PCM technology has made rapid progress in a short time, having passed older technologies in terms of both sophisticated demonstrations…
▽ More
We survey the current state of phase change memory (PCM), a non-volatile solid-state memory technology built around the large electrical contrast between the highly-resistive amorphous and highly-conductive crystalline states in so-called phase change materials. PCM technology has made rapid progress in a short time, having passed older technologies in terms of both sophisticated demonstrations of scaling to small device dimensions, as well as integrated large-array demonstrators with impressive retention, endurance, performance and yield characteristics.
We introduce the physics behind PCM technology, assess how its characteristics match up with various potential applications across the memory-storage hierarchy, and discuss its strengths including scalability and rapid switching speed. We then address challenges for the technology, including the design of PCM cells for low RESET current, the need to control device-to-device variability, and undesirable changes in the phase change material that can be induced by the fabrication procedure. We then turn to issues related to operation of PCM devices, including retention, device-to-device thermal crosstalk, endurance, and bias-polarity effects. Several factors that can be expected to enhance PCM in the future are addressed, including Multi-Level Cell technology for PCM (which offers higher density through the use of intermediate resistance states), the role of coding, and possible routes to an ultra-high density PCM technology.
△ Less
Submitted 28 March, 2010; v1 submitted 7 January, 2010;
originally announced January 2010.
