Periodic Reporting for period 4 - PROJESTOR (PROJECTED MEMRISTOR: A nanoscale device for cognitive computing)
Período documentado: 2021-01-01 hasta 2021-06-30
As complementary metal–oxide–semiconductor (CMOS) scaling reaches its technological limits, a radical departure from traditional von Neumann systems, which involve separate processing and memory units, is needed in order to extend the performance of today’s computers substantially. In-memory computing is a promising approach in which nanoscale memory devices, organized in a computational memory unit, are used for both processing and memory. The essential idea is to exploit the physical attributes and state dynamics of these memory devices to compute in place without the need to shuttle the data back and forth between the memory and processing units. The critical element in these novel computing systems is a very-high-density, low-power, variable-state, programmable non-volatile memory device. The overall objective of PROJESTOR was to explore such a memory device as well as to investigate ways in which we can use such memory devices for computational purposes.
PROJESTOR made critical contributions to both the above-stated goals. On the device front, the project helped establish phase-change memory and in particular projected phase-change memory as a powerful memory device concept for in-memory computing. The key scientific accomplishments in the device front include identification of Antimony as a single elemental phase change memory, significant understanding of structural relaxation in melt-quenched phase-change materials and the demonstration of 8-bit equivalent precision computation using projected phase-change memory. On the application side, a key breakthrough was the concept of mixed-precision in-memory computing which helped expand the application space of analog in-memory computing. We could also identify new applications such as hyperdimensional computing and in-memory database query and experimentally verify using phase-change memory devices.
PROJESTOR made critical contributions to both the above-stated goals. On the device front, the project helped establish phase-change memory and in particular projected phase-change memory as a powerful memory device concept for in-memory computing. The key scientific accomplishments in the device front include identification of Antimony as a single elemental phase change memory, significant understanding of structural relaxation in melt-quenched phase-change materials and the demonstration of 8-bit equivalent precision computation using projected phase-change memory. On the application side, a key breakthrough was the concept of mixed-precision in-memory computing which helped expand the application space of analog in-memory computing. We could also identify new applications such as hyperdimensional computing and in-memory database query and experimentally verify using phase-change memory devices.
The first part of the project aimed at developing memory devices for in-memory computing. The goal was to explore the concept of projected memory devices and in particular projected phase-change memory devices where the read and write operations are decoupled by exploiting the nonlinear electrical transport in amorphous phase change materials.
As part of this study, we explored various phase-change materials and projection materials. The most important invention in this front was that of single elemental phase-change materials. We showed how the simplest material imaginable, a single element (in this case, antimony), can become a valid alternative when confined in extremely small volumes. This work appeared on the cover of Nature Materials (August 2018 issue) [1]. We also developed the most comprehensive understanding of resistance drift in phase-change materials, the key attribute that we are trying to counter through the concept of projection. This work was published in Advanced Electronics Materials [2]. More recently, we could experimentally measure the onset of structural relaxation in the melt-quenched amorphous phase [3]. We also established a comprehensive understanding of projected phase change memory [4] and also demonstrated the ability to achieve 8-bit equivalent precision arithmetic operations [5]. The latter work was selected as a highlight paper at IEDM.
In the second part of the project, we explored in-memory computing. We showed several applications where we could exploit the physical attributes and state dynamics of phase-change memory devices to perform computation in place. In one application, we showed how we could perform compressed sensing and recovery. This work was published in IEEE Transactions on Electron Devices [6]. We also demonstrated yet another application where we exploit the crystallization dynamics of phase-change memory devices to solve an unsupervised learning algorithm entirely in the memory [7]. A significant challenge for computational memory is the lack of precision associated with it. We proposed a new concept called mixed precision in-memory computing to address this. We used this concept to solve systems of linear equations and this work appeared in Nature Electronics [8]. The mixed-precision in-memory computing concept was also extended to tackle the challenging problem of training deep neural networks [9]. We also developed a multi-PCM device architecture to counter some of the short comings of in-memory computing [10]. We also explored ways to train networks to be resilient to analog in-memory computing [11]. We also explored other applications such as in-memory hyperdimensional computing [12]. Finally, we wrote a couple of tutorials and review articles that have gained a lot of attention in recent years [13,14]
[1] Salinga, Kersting et al., Nature Materials, 2018 (Cover story)
[2] Le Gallo et al., Adv. Electr. Materials, 2018
[3] Kersting et al., Adv. Func. Mat., 2021
[4] Kersting et al., Scientific Reports, 2020
[5] Giannopoulos et al., Proc. IEDM, 2018 (Highlight paper)
[6] Le Gallo et al., IEEE TED., 2018
[7] Sebastian et al., Nature Comm., 2017
[8] Le Gallo et al., Nature Electronics, 2018
[9] Nandakumar et al., Front. Neuroscience, 2020
[10] Boybat et al., Nature Comm., 2018
[11] Joshi et al., Nature Comm., 2020
[12] Karunaratne et al., Nature Electronics, 2020 (Cover story)
[13] Sebastian et al, J. Appl. Phys., 2018
[14] Sebastian et al., Nature Nanotech., 2020 (Cover story)
As part of this study, we explored various phase-change materials and projection materials. The most important invention in this front was that of single elemental phase-change materials. We showed how the simplest material imaginable, a single element (in this case, antimony), can become a valid alternative when confined in extremely small volumes. This work appeared on the cover of Nature Materials (August 2018 issue) [1]. We also developed the most comprehensive understanding of resistance drift in phase-change materials, the key attribute that we are trying to counter through the concept of projection. This work was published in Advanced Electronics Materials [2]. More recently, we could experimentally measure the onset of structural relaxation in the melt-quenched amorphous phase [3]. We also established a comprehensive understanding of projected phase change memory [4] and also demonstrated the ability to achieve 8-bit equivalent precision arithmetic operations [5]. The latter work was selected as a highlight paper at IEDM.
In the second part of the project, we explored in-memory computing. We showed several applications where we could exploit the physical attributes and state dynamics of phase-change memory devices to perform computation in place. In one application, we showed how we could perform compressed sensing and recovery. This work was published in IEEE Transactions on Electron Devices [6]. We also demonstrated yet another application where we exploit the crystallization dynamics of phase-change memory devices to solve an unsupervised learning algorithm entirely in the memory [7]. A significant challenge for computational memory is the lack of precision associated with it. We proposed a new concept called mixed precision in-memory computing to address this. We used this concept to solve systems of linear equations and this work appeared in Nature Electronics [8]. The mixed-precision in-memory computing concept was also extended to tackle the challenging problem of training deep neural networks [9]. We also developed a multi-PCM device architecture to counter some of the short comings of in-memory computing [10]. We also explored ways to train networks to be resilient to analog in-memory computing [11]. We also explored other applications such as in-memory hyperdimensional computing [12]. Finally, we wrote a couple of tutorials and review articles that have gained a lot of attention in recent years [13,14]
[1] Salinga, Kersting et al., Nature Materials, 2018 (Cover story)
[2] Le Gallo et al., Adv. Electr. Materials, 2018
[3] Kersting et al., Adv. Func. Mat., 2021
[4] Kersting et al., Scientific Reports, 2020
[5] Giannopoulos et al., Proc. IEDM, 2018 (Highlight paper)
[6] Le Gallo et al., IEEE TED., 2018
[7] Sebastian et al., Nature Comm., 2017
[8] Le Gallo et al., Nature Electronics, 2018
[9] Nandakumar et al., Front. Neuroscience, 2020
[10] Boybat et al., Nature Comm., 2018
[11] Joshi et al., Nature Comm., 2020
[12] Karunaratne et al., Nature Electronics, 2020 (Cover story)
[13] Sebastian et al, J. Appl. Phys., 2018
[14] Sebastian et al., Nature Nanotech., 2020 (Cover story)
During the course of the project, we made some significant progress beyond the state of the art in both the development of new memory devices as well as their application in computing.
In the memory device front, the exploration of singe elemental phase-change materials is a paradigm shift in this field given that up to now, much of the effort was on fine tuning the material composition. With our work we have shown that, it is essential to focus on nanoscale size effects instead. We also explored projected PCM devices with single elemental phase-change materials as well projected phase change devices based on a mushroom-type device geometry.
We also gained a deep understanding of structural relaxation in melt-quenched amorphous phase-change materials. We established the collective relaxation model that quantitatively captures structural relaxation. More recently, we showed that the collective relaxation model could also explain the onset of relaxation that we experimentally measured.
In the computing front, we developed numerous applications for in-memory computing from compressed sensing and temporal correlation detection to solving systems of linear equations. We also showed that we can achieve in-memory scalar operations with projected PCM devices with an equivalent precision of 8-bit fixed point arithmetic and consuming less than 10fJ. This is the first time such high precision is reported using in-memory computing. In the second half of the project, we made critical contributions to in-memory deep learning. We established ways to achieve software-equivalent accuracy training and inference using in-memory computing. We also explored machine learning algorithms that transcend deep learning such as hyperdimensional computing.
In the memory device front, the exploration of singe elemental phase-change materials is a paradigm shift in this field given that up to now, much of the effort was on fine tuning the material composition. With our work we have shown that, it is essential to focus on nanoscale size effects instead. We also explored projected PCM devices with single elemental phase-change materials as well projected phase change devices based on a mushroom-type device geometry.
We also gained a deep understanding of structural relaxation in melt-quenched amorphous phase-change materials. We established the collective relaxation model that quantitatively captures structural relaxation. More recently, we showed that the collective relaxation model could also explain the onset of relaxation that we experimentally measured.
In the computing front, we developed numerous applications for in-memory computing from compressed sensing and temporal correlation detection to solving systems of linear equations. We also showed that we can achieve in-memory scalar operations with projected PCM devices with an equivalent precision of 8-bit fixed point arithmetic and consuming less than 10fJ. This is the first time such high precision is reported using in-memory computing. In the second half of the project, we made critical contributions to in-memory deep learning. We established ways to achieve software-equivalent accuracy training and inference using in-memory computing. We also explored machine learning algorithms that transcend deep learning such as hyperdimensional computing.