Fast Memory Research Articles

Fast and Low-GPU-memory abdomen CT organ segmentation: The FLARE challenge.

Automatic segmentation of abdominal organs in CT scans plays an important role in clinical practice. However, most existing benchmarks and datasets only focus on segmentation accuracy, while the model efficiency and its accuracy on the testing cases from different medical centers have not been evaluated. To comprehensively benchmark abdominal organ segmentation methods, we organized the first Fast and Low GPU memory Abdominal oRgan sEgmentation (FLARE) challenge, where the segmentation methods were encouraged to achieve high accuracy on the testing cases from different medical centers, fast inference speed, and low GPU memory consumption, simultaneously. The winning method surpassed the existing state-of-the-art method, achieving a 19× faster inference speed and reducing the GPU memory consumption by 60% with comparable accuracy. We provide a summary of the top methods, make their code and Docker containers publicly available, and give practical suggestions on building accurate and efficient abdominal organ segmentation models. The FLARE challenge remains open for future submissions through a live platform for benchmarking further methodology developments at https://flare.grand-challenge.org/.

NORM: An FPGA-based Non-volatile Memory Emulation Framework for Intermittent Computing

Today’s intermittent computing systems operate by relying only on harvested energy accumulated in their tiny energy reservoirs, typically capacitors. An intermittent device dies due to a power failure when there is no energy in its capacitor and boots again when the harvested energy is sufficient to power its hardware components. Power failures prevent the forward progress of computation due to the frequent loss of computational state. To remedy this problem, intermittent computing systems comprise built-in fast non-volatile memories with high write endurance to store information that persists despite frequent power failures. However, the lack of design tools makes fast-prototyping these systems difficult. Even though FPGAs are common platforms for fast prototyping and behavioral verification of continuously powered architectures, they do not target prototyping intermittent computing systems. This article introduces a new FPGA-based framework, named NORM ( N on-volatile mem OR y e M ulator), to emulate and verify the behavior of any intermittent computing system that exploits fast non-volatile memories. Our evaluation showed that NORM can be used to emulate and validate FeRAM-based transiently powered hardware architectures successfully.

Low-Latency Nested Decoding for Short Generalized Integrated Interleaved BCH Codes

Generalized integrated interleaved (GII) codes nest short BCH sub-codewords to form more powerful BCH codewords. They can potentially achieve hyper-speed decoding with excellent error-correction capability. In particular, short GII-BCH codes are among the best candidates for the new fast storage class memories (SCMs). Miscorrections severely degrade the performance of short GII-BCH codes. Although they were effectively mitigated in previous designs, the involved repeated Chien search and higher-order syndrome computation cause long latency. This brief proposes efficient and low-latency nested decoding schemes for short GII-BCH codes. A strategy is developed to select sub-words for further nested decoding to mitigate miscorrections by keeping track of the error locator polynomials, instead of waiting for the lengthy Chien search. Formulas are also derived to estimate the effects on the error-correcting performance. Besides, a low-complexity linear feedback shift register (LFSR) architecture is developed to accelerate the higher-order nested syndrome computation. For an example GII-BCH code targeting at SCMs, the proposed design reduces the worst-case nested decoding latency by 26% with 8.5% area overhead and negligible performance loss compared to prior methods.

Temperament and children’s accuracy and attention during word learning

When hearing a novel word, children typically rule out familiar objects and assume a speaker is referring to a novel object. This strategy is known as fast mapping, and young children use this with a high degree of accuracy. However, not all children engage in fast mapping to the same extent and temperament can play a role. Shyness is associated with poorer fast mapping and less attention to target objects, which is associated with poorer retention (Hilton et al., 2019; Hilton & Westermann, 2017). We further investigated the relationship between temperament and fast mapping by presenting 2.5-year-old children with 8 familiar target fast mapping trials and 4 novel target trials presented twice. We considered two temperamental dimensions: approachability due to its similarity to shyness; and reactivity, which could predict children’s capacity to engage during fast mapping. We found an association between approachability and fast mapping accuracy the second time children fast-mapped novel targets, and approachability was associated with greater retention accuracy. Reactivity predicted proportions of target looking during fast mapping with less reactive temperament scores associated with greater focus on targets. This provides support for a relationship between two dimensions of temperament and fast mapping and retention. Approachability may be associated with a further opportunity to fast map and memory for novel words, and/or how willing children are to guess the targets. Reactivity may be associated with the capacity to focus during word learning situations. Different aspects of temperament could have implications for children’s capacity to disambiguate and learn words.

Data Plane Cooperative Caching With Dependencies

Caching is at the core of most modern communication systems, where caches are used to store content and traffic classification rules. While network components can leverage caching in a cooperative manner, one important aspect of such systems concerns possible dependencies among stored items. A major use case of such dependencies appears in rule placement across software-defined networks (SDNs). Despite the tremendous success of SDNs in datacenters, their wide adoption still poses a key challenge: the packet-forwarding rules in switches require fast and power-hungry memories. Rule tables, which serve as caches, are of limited size in cheap and energy-constrained devices, motivating novel solutions to achieve high hit rates. We leverage device connectivity in the fast data plane, where delays are in the order of few milliseconds, and propose multiple switches to work together to avoid accessing the control plane, where delays are orders of magnitude greater. As a low priority rule in a cache entails caching higher priority rules, we pose the problem of cooperative caching with dependencies. We provide models and algorithms accounting for dependencies among rules implied by existing switch memory types, and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">lay the foundations of cooperative caching with dependencies</i> .

Smart, Fast, and Low Memory Beam-Steering Antenna Configurations for 5G and Future Wireless Systems

Smart Antennas are important to provide mobility support for many enhanced 5G and future wireless applications and services, such as energy harvesting, virtual reality, Voice over 5G (Vo5G), connected vehicles, Machine-to-Machine Communication (M2M), and Internet of Things (IoT). Smart antenna technology enables us to reduce interference and multipath problems and increase the quality in communication signals. This paper presents a number of nonlinear configurations of dipole arrays for forming a single beam in any desired direction. We propose three, four, six, and eight-element array structures to perform this single beam-steering functionality. The proposed array configurations with multiple axes of symmetry (in the azimuthal plane) decrease the computational repetitions in optimizing respective weight factors for beam-steering. The optimized weight factors are obtained through the Least Mean Square (LMS) method. MATLABTM is used to calculate optimized weight factors as well as to determine the resulting radiation patterns. Since antennas are bidirectional elements, beamforming in one direction means that the antenna will also have high receiving gain in that direction. Performances of differently configured models are compared in terms of their directivity, sidelobe reduction, and computational complexities for beam-steering.

Controlling a solar receiver with multiple thermochemical reactors for hydrogen production by an LSTM neural network based cascade controller

Solar reactors for batch processes have a varying heat demand over time. The receiver of large-scale solar chemical plants consists of several of these reactors to quasi-continuously use the solar resource. For this, the heliostat field needs to provide a non-uniform time-varying flux density distribution. As this differs from the heliostat field control of a typical CSP plant, it challenges its control. Here, we propose a cascade controller for a coupled system of heliostat field and receiver consisting of 19 thermochemical batch reactors for hydrogen generation. In the receiver controller, a fast long short-term memory (LSTM) neural network model predicts the future temperatures and the ceria reduction extends in the reactors as a function of the flux density. It is trained with data from a thermochemical reactor model. The trained LSTM neural network is embedded in a hybrid automaton to model the continuous batch process states. Discrete state changes are made when certain conditions are reached. In this way, the receiver controller determines flux density setpoints leading to high efficiencies for each reactor. The heliostat field controller applies aim point optimization to provide flux densities close to these setpoints. With this cascade control, reactor array efficiencies of 79% are obtained simulatively. Moreover, the individual reactors operate within their material limits and come close to their optimal efficiency. In addition, it is found that secondary concentrators complicate the control and decrease the reactor array efficiency. However, they still increase the overall efficiency by 51.3% due to a significantly higher optical efficiency.

Mesoscopic superconducting memory based on bistable magnetic textures

With the ever-increasing energy need to process big data, the realization of low-power computing technologies, such as superconducting logic and memories, has become a pressing issue. Developing fast and non-volatile superconducting memory elements, however, remains a challenge. Superconductor-ferromagnet hybrid devices offer a promising solution, as they combine ultra-fast manipulation of spins with dissipationless readout. Here, we present a new type of non-volatile Josephson junction memory that utilizes the bistable magnetic texture of a single mesoscopic ferromagnet. We use micromagnetic simulations to design an ellipse-shaped planar junction structured from a Nb/Co bilayer. The ellipse can be prepared as uniformly magnetized or as a pair of vortices at zero applied field. The two states yield considerably different critical currents, enabling reliable electrical readout of the element. We describe the mechanism used to control the critical current by applying numerical calculations to quantify the local stray field from the ferromagnet, which shifts the superconducting interference pattern. By combining micromagnetic modeling with bistable spin-textured junctions, our approach presents a novel route towards realizing superconducting memory applications.

Polarization of Bi2Se3 thin film toward non-volatile memory applications

In recent years, topological insulators have drawn growing interest as a unique electronic state of matter toward quantum information technology. Despite the logic devices with magnetization switching through spin–orbit torque or the topological magneto-electric effect, realizing memory devices based on topological insulators has been urged in quantum computing applications. In this work, we report the design and fabrication of a non-volatile memory device that employs polarization of Bi2Se3 thin films achieving fast memory speed, sufficient memory window, and good reliability. The Bi2Se3 film polarizes under an external electrical field with charges accumulated on the top and bottom surfaces separating the electrons and holes. Such polarization is much faster than the carrier tunneling in conventional floating-gate flash memory and ferroelectric-based memory devices. In addition, good memory retention and endurance properties have also been obtained, showing great potential in high-performance memory application in future topological insulator-involved information technology.

Comparison of Apriori, Apriori-TID and FP-Growth Algorithms in Market Basket Analysis at Grocery Stores

Market Basket Analysis is an analysis of consumer behavior specifically from a certain group/group. Market Basket Analysis is generally used as a starting point for seeking knowledge from a data transaction when we do not know what specific pattern we are looking for. Market Basket Analysis in this study is applied to the search for patterns of purchasing groceries at grocery stores and then analyzed by season. This study aims to compare the Apriori, Apriori TID and FP-Growth methods in determining consumer transaction behavior and calculating the quantity of consumer transactions in several seasons based on data obtained from the Market Basket Analysis database. In the results of this study, it is known that FP-Growth has the best performance among the other two algorithms, but uses more memory than other algorithms. The Apriori-TID algorithm uses lighter and faster memory than the Apriori Algorithm

Zero-Field Nucleation and Fast Motion of Skyrmions Induced by Nanosecond Current Pulses in a Ferrimagnetic Thin Film.

Skyrmion racetrack memories are highly attractive for next-generation data storage technologies. Skyrmions are noncollinear spin textures stabilized by chiral interactions. To achieve a fast-operating memory device, it is critical to move skyrmions at high speeds. The skyrmion dynamics induced by spin-orbit torques (SOTs) in the commonly studied ferromagnetic films is hindered by strong pinning effects and a large skyrmion Hall effect causing deflection of the skyrmion toward the racetrack edge, which can lead to information loss. Here, we investigate the current-induced nucleation and motion of skyrmions in ferrimagnetic Pt/CoGd/(W or Ta) thin films. We first reveal field-free skyrmion nucleation mediated by Joule heating. We then achieve fast skyrmion motion driven by SOTs with velocities as high as 610 m s-1 and a small skyrmion Hall angle |θSkHE| ≲ 3°. Our results show that ferrimagnets are better candidates for fast skyrmion-based memory devices with low risk of information loss.

Enhancing data locality of the conjugate gradient method for high-order matrix-free finite-element implementations

This work investigates a variant of the conjugate gradient (CG) method and embeds it into the context of high-order finite-element schemes with fast matrix-free operator evaluation and cheap preconditioners like the matrix diagonal. Relying on a data-dependency analysis and appropriate enumeration of degrees of freedom, we interleave the vector updates and inner products in a CG iteration with the matrix-vector product with only minor organizational overhead. As a result, around 90% of the vector entries of the three active vectors of the CG method are transferred from slow RAM memory exactly once per iteration, with all additional access hitting fast cache memory. Node-level performance analyses and scaling studies on up to 147k cores show that the CG method with the proposed performance optimizations is around two times faster than a standard CG solver as well as optimized pipelined CG and s-step CG methods for large sizes that exceed processor caches, and provides similar performance near the strong scaling limit.

Wet Chemical Recess Etching of Ge2Sb2Te5 for 3D PCRAM Memory Applications

Memory cells comprising a Phase Change Material (PCM) are the building blocks of fast and non-volatile memory devices called Phase Change Random Access Memory (PCRAM) [1-3]. The working principle of this memory involves data retention in the form of a phase (amorphous or crystalline) and the set and reset can be done by Joule heating to induce an amorphous-to-crystalline or crystalline-to-amorphous transition respectively. Some chalcogenide materials experience this thermally driven phase change, GeSbTe (GST) being one of those alloys extensively studied. GST has also been adopted for the fabrication of the 1st generation X-point memory [4] and might be adopted in a 2nd generation X-point memory of a four-layer PCM structure [5]. However, this 2D scaling or stacking of PCRAM is limited by cost and therefore the development of 3D architectures is envisaged for decreasing the cost/bit [6].A key requirement for the fabrication of this 3D architecture is the conformal deposition and etch-back of GST. Dry plasma etching might be limited to anisotropic recess while isotropic lateral recess is needed. Therefore, wet isotropic etching might be the process of choice. A few chemical solutions have been proposed in previous studies. Cheng et al. showed that GST could be etched in HNO3 but with a very high etch rate and with an unwanted surface composition change due to different oxidation and dissolution rates of the metalloids [7]. Wang et al. demonstrated that basic wet etching solutions led to a slower etch rate and a much smoother surface compared to acidic wet etching solutions [8]. Deng et al. showed a switch in the etch rate order between crystalline and amorphous GST depending on the H2O2 concentration in TMAH [9].In this work, we present a controllable partial recess solution that leaves the GST surface smooth after recess. Wet recess of amorphous and crystalline blanket films, as well as patterned samples, was initially explored using the commodity chemistries Ammonium Peroxide Mixture (APM) and (Hydrochloric Peroxide Mixture) HPM. The etching of GST in HPM as a function of the H2O2 concentration was monitored by ICPMS and showed a well-controlled etch rate. However, some shortcomings of these H2O2-containing solutions, like roughness and selectivity, lead to a change of oxidizing agent from H2O2 to O3. In the O3-containing solutions, the selectivity towards Al2O3, SiO2, and TiN could be secured. The impact of the dissolved O3 concentration on surface roughness and etch rate as well as the uniformity of this wet etching process were assessed on a single wafer tool. Finally, the bulk and surface GST composition and oxidation post-recess were verified through XPS and ERD.REFERENCES:[1] D. Loke et al., “Breaking the Speed Limits of Phase-Change Memory.” Science, 2012, 336, 6088, 1566.[2] K. Ding et al., “Recipe for ultrafast and persistent phase-change memory materials.” NPG Asia Mater 12, 63, 2020.[3] F. Rao et al., “Reducing the stochasticity of crystal nucleation to enable sub-nanosecond memory writing.” Science, 2017, 358, 6369, 1423.[4] [internet] https://www.techinsights.com/blog/intel-3d-xpoint-memory-die-removed-intel-optanetm-pcm-phase-change-memory [5] [internet] https://www.techinsights.com/blog/memory/intels-2nd-generation-xpoint-memory [6] [internet] https://www.imec-int.com/en/imec-magazine/imec-magazine-october-2017/in-pursuit-of-high-density-storage-class-memory [7] H.Y. Cheng et al., “Wet-Etching Characteristics of Ge2Sb2Te5 Thin Films for Phase-Change Memory.” IEEE Trans. Magn., 41, 2, 2005.[8] L. Wang et al., “Basic Wet-Etching Solutions for Ge2Sb2Te5 Phase Change Material.” J. Electrochem. Soc., 157, H470, 2010.[9] C. Deng et al., “XPS study on the selective wet etching mechanism of GeSbTe phase change thin films with TMAH.” Proc. of SPIE, 8782, 87820N, 2012.

Two Manchester Computer Milestones

On June 21, 2022 in Manchester, England, the IEEE unveiled two bronze Milestone plaques in recognition of two historic computing developments. The first honored the Manchester University Baby computer and its derivatives, 1948–1951. The second plaque marked the Atlas computer and the invention of virtual memory, 1957– 1962. Both projects benefitted from a significant long-term collaboration between the University and industry (Ferranti Ltd.). The formal citations are as follows: At this site, on June 21 1948, the “Baby” became the first computer to execute a program stored in addressable read-write electronic memory. “Baby” validated Williams-Kilburn Tube random-access memories, later widely used, and led to the 1949 Manchester Mark I which pioneered index registers. In February 1951, Ferranti Ltd.’s commercial derivative became the first electronic computer marketed as a standard product delivered to a customer. The Atlas computer was designed and built in this building by Tom Kilburn and a joint team of the University of Manchester and Ferranti Ltd. The most significant new feature of Atlas was the invention of virtual memory, allowing memories of different speeds and capacities to act as a single large fast memory separately available to multiple users. Virtual memory became a standard feature of general-purpose computers.

Privacy Protection and Intrusion Detection System of Wireless Sensor Network Based on Artificial Neural Network.

With the increasing openness and development of network technology, the network based on the wireless sensor network system has increasingly become an important tool for human social life and production, but it also brings some network security problems. Among them, focus is on network privacy disclosure and foreign intrusion and the research of intrusion detection and privacy protection has increasingly become an important topic of network security. This paper deeply studies the wireless sensor network system based on neural network on the basis of traditional privacy protection and intrusion detection system. Firstly, it applies particle swarm optimization algorithm and constructs a wireless sensor network intrusion detection system based on particle swarm optimization algorithm. The system includes important modules such as data extraction, data analysis, data feedback, and auxiliary decision-making. Compared with other algorithms, particle swarm optimization algorithm does not rely on problem information. It mainly uses real numbers to solve, so the algorithm has strong universality. At the same time, its corresponding principle is simple and easy to implement and less parameters need to be adjusted. Compared with other algorithms, particle swarm optimization algorithm has fast convergence speed and little memory requirement for the computer. At the same time, this paper uses the leap of particle swarm optimization algorithm to make it easier to find the global optimal solution. At the corresponding level of wireless sensor network privacy protection, based on the original data aggregation privacy protection scheme, this paper proposes a privacy protection scheme based on polynomial regression and a user privacy protection scheme based on the same state encryption, which further improves the security of privacy protection and facilitates the management of information. To realize the integrity of user privacy information protection, this paper realizes the decryption of data based on the correlation between binary metadata and compares the corresponding decrypted data with aggregated data, so as to complete the integrity of privacy data protection. The experimental results show that the binary metadata correlation decryption method proposed in this paper and the introduction of the corresponding particle swarm optimization algorithm improve the stability of the system by about 10%, the corresponding system security by a positive proportion, and the integrity of private data by about 10%; therefore, the algorithm proposed in this paper has obvious advantages.

Current-induced self-switching of perpendicular magnetization in CoPt single layer

All-electric switching of perpendicular magnetization is a prerequisite for the integration of fast, high-density, and low-power magnetic memories and magnetic logic devices into electric circuits. To date, the field-free spin-orbit torque (SOT) switching of perpendicular magnetization has been observed in SOT bilayer and trilayer systems through various asymmetric designs, which mainly aim to break the mirror symmetry. Here, we report that the perpendicular magnetization of CoxPt100-x single layers within a special composition range (20 < x < 56) can be deterministically switched by electrical current in the absence of external magnetic field. Specifically, the Co30Pt70 shows the largest out-of-plane effective field efficiency and best switching performance. We demonstrate that this unique property arises from the cooperation of two structural mechanisms: the low crystal symmetry property at the Co platelet/Pt interfaces and the composition gradient along the thickness direction. Compared with that in bilayers or trilayers, the field-free switching in CoxPt100-x single layer greatly simplifies the SOT structure and avoids additional asymmetric designs.

Mo-based perpendicularly magnetized thin films with low damping for fast and low-power consumption magnetic memory

Mo-based perpendicularly magnetized thin films with low damping for fast and low-power consumption magnetic memory

Magnetization reversal properties and magnetostatic interactions of disk to rod-shaped FeNi layers separated by ultra-thin Cu layers

From fast magnetic memories with low-power consumption to recording media with high densities, realizing the magnetization reversal and interaction of magnetic layers would allow for manipulating the ultimate properties. Here, we use a pulsed electrochemical deposition technique in porous alumina templates (50 nm in pore diameter) to fabricate arrays of nanowires, consisting of FeNi layers (26−227 nm in thickness) with disk to rod-shaped morphologies separated by ultra-thin (3 nm) Cu layers. By acquiring hysteresis curves and first-order reversal curves (FORCs) of the multilayer nanowire arrays, we comprehensively investigate magnetization reversal properties and magnetostatic interactions of the layers at different field angles (0° ≤ θ ≤ 90°). These involve the extraction of several parameters, including hysteresis curve coercivity (H c Hyst ), FORC coercivity (H c FORC ), interaction field distribution width (ΔH u ), and irreversible fraction of magnetization (IF m ) as a function of θ. We find relatively constant and continuously decreasing trends of H c Hyst when 0° ≤ θ ≤ 45°, and 45° < θ ≤ 90°, respectively. Meanwhile, angular dependence of H c FORC and IF m shows continuously increasing and decreasing trends, irrespective of the FeNi layer morphology. Our FORC results indicate the magnetization reversal properties of the FeNi/Cu nanowires are accompanied with vortex domain wall and single vortex modes, especially at high field angles. The rod-shaped layers also induce maximum ΔH u during the reversal process, owing to enhancements in both magnetizing and demagnetizing-type magnetostatic interactions.

An Enterprise-Grade Open-Source Data Reduction Architecture for All-Flash Storage Systems

All-flash storage (AFS) systems have become an essential infrastructure component to support enterprise applications, where sub-millisecond latency and very high throughput are required. Nevertheless, the price per capacity ofsolid-state drives (SSDs) is relatively high, which has encouraged system architects to adoptdata reduction techniques, mainlydeduplication andcompression, in enterprise storage solutions. To provide higher reliability and performance, SSDs are typically grouped usingredundant array of independent disk (RAID) configurations. Data reduction on top of RAID arrays, however, adds I/O overheads and also complicates the I/O patterns redirected to the underlying backend SSDs, which invalidates the best-practice configurations used in AFS. Unfortunately, existing works on the performance of data reduction do not consider its interaction and I/O overheads with other enterprise storage components including SSD arrays and RAID controllers. In this paper, using a real setup with enterprise-grade components and based on the open-source data reduction module RedHat VDO, we reveal novel observations on the performance gap between the state-of-the-art and the optimal all-flash storage stack with integrated data reduction. We therefore explore the I/O patterns at the storage entry point and compare them with those at the disk subsystem. Our analysis shows a significant amount of I/O overheads for guaranteeing consistency and avoiding data loss through data journaling, frequent small-sized metadata updates, and duplicate content verification. We accompany these observations with cross-layer optimizations to enhance the performance of AFS, which range from deriving new optimal hardware RAID configurations up to introducing changes to the enterprise storage stack. By analyzing the characteristics of I/O types and their overheads, we propose three techniques: (a) application-aware lazy persistence, (b) a fast, read-only I/O cache for duplicate verification, and (c) disaggregation of block maps and data by offloading block maps to a very fast persistent memory device. By consolidating all proposed optimizations and implementing them in an enterprise AFS, we show 1.3× to 12.5× speedup over the baseline AFS with 90% data reduction, and from 7.8× up to 57× performance/cost improvement over an optimized AFS (with no data reduction) running applications ranging from 100% read-only to 100% write-only accesses.

Spatially dependent optical bistability

The effect of the orbital angular momentum (OAM) of light beams on the optical bistability (OB) phenomenon has been investigated in a three-level quantum system. In a first step, we consider a Λ -type configuration in which the coupling field is assumed as a Laguerre–Gaussian (LG) beam. It is shown that the OB strength spatially depends on the intensity profile as well as the topological charge (TC) of the light beam. In a second scheme, we apply a strong microwave field to the lower level transition and obtain an analytical expression for the OB behavior of the probe field in the presence of an LG field in the multi-photon resonance condition. We analytically obtain a generalized input–output bistability relation for the systems containing the LG beam. More complicated phase-dependent OB profiles can be obtained for higher order LG modes due to the OAM transfer and classical interference between the generated field and applied fields. The results of this paper may provide a potential application of OB in generating simultaneously reconfigurable sets of various optical logic gates in the same structure, as well as in enhancing the capacity of fast volatile bit-level optical memories.