Area Budget Research Articles

Euro Area Macroeconomics. Where Do We Stand 20 Years Later?

For almost 20 years, euro area countries have been sharing a single currency. The drawbacks of the euro area framework were highlighted by the widening of imbalances prior to the 2007 financial crisis, and thereafter by the huge impact of the financial crisis, the public debt crisis in Southern European countries, and the Great Recession. Prior to and after the crisis, EU institutions and Member States (MS) have not been able to implement either a common economic strategy or satisfactory economic policy coordination.This led neither to a bursting of the euro area, nor to a substantial change in its functioning. Euro area institutions were adapted, through the European Stability Mechanism, the Fiscal Treaty, the “first semester”, the European Central Bank’s support to MS, and the banking union. These adaptations were painful.In mid-2018, the economic situation had clearly improved at the euro area level. However, the following question remains unsolved: can the functioning of the euro area be improved, accounting for divergent situations, interests and views in MS?The paper recalls proposals from EU institutions and from MS. We present and discuss a number of proposals made by economists to improve the euro area policy framework: relying on financial markets to control domestic economic policies, introducing a euro area budget and a minister of finance, moving towards a federal EU with increased democracy, and last, improving economic policy coordination.

Hetro8T: power and area efficient approximate heterogeneous 8T SRAM for H.264 video decoder

Wide-spread availability of high-speed INTERNET and rapid increase of smart-phone users have significantly increased online video surfing. Video decoders like H.264/H.265/MPEG consume a significant amount of power in Static Random Access Memory (SRAM) buffers. In this study, the authors propose a 1 kb (32 × 32) heterogeneous 8T SRAM architectures with (2-lower order bits) and without truncation for H.264 video decoder. They have used heterogeneous sized SRAM design and bit-truncation techniques are used simultaneously to obtain low power memory design for the H.264 video decoder. They show that the proposed approximate memory used for H.264 video decoder provide high video quality even at low power and low area budget of 0.3 µW/pixel and 5.2 µm 2 /pixel, respectively, at 0.5 V and 20 MHz in UMC 28 nm CMOS technology. The proposed memory architecture is compared with existing approximate memories such as heterogeneous 6T, hybrid 8T/6T, all-identical 6T, and all-identical 8T SRAM memory. The results show that proposed memory architectures perform cumulatively better than existing techniques in terms of dynamic power, leakage power, and area.

Carbon dioxide emissions evaluations and mitigations in the building and traffic sectors in Taichung metropolitan area, Taiwan

Since the beginning of the industrial revolution, the atmospheric CO2 level has been continuously increasing because of the overall energy use in urban areas that generates excessively high levels of CO2 emissions. This study reviewed the relevant literature, then adopted the Taichung metropolitan area as the research target and assessed carbon emissions with respect to buildings, traffic, and carbon sinks. The overall carbon budget of the metropolitan area was mapped following a statistical analysis of the numerical data and urban space information. The results of this study indicated that the annual average urban carbon emissions consist of 67.6% building carbon emissions, 30.5% traffic carbon emissions, and −1.9% carbon sink absorption. In this study, a multiple regression model was used to calculate the floor area of each building. This study also determined that densely populated areas emitted higher levels of carbon than less populated areas. For every square meter of total floor area, 16.51 tCO2/m2·yr of carbon were emitted from buildings every year. Recommended policies for the city government to implement in the future were organized and used to establish three simulation scenarios of the various implementation stages. The results indicated that adjusting the floor area ratio of buildings is the optimal carbon reduction approach, achieving a reduction of at most 620,363 tons of carbon per year, and multi-scale carbon reduction hotspots were mapped out. In addition, the high-resolution grid was used to present the multi-scale carbon budget results, which helps government agencies to formulate follow-up priority carbon reduction strategies and urban carbon neutral policies.

Baffin Bay sea ice inflow and outflow: 1978–1979 to 2016–2017

Abstract. Baffin Bay serves as a huge reservoir of sea ice which would provide the solid freshwater sources to the seas downstream. By employing satellite-derived sea ice motion and concentration fields, we obtain a nearly 40-year-long record (1978–1979 to 2016–2017) of the sea ice area flux through key fluxgates of Baffin Bay. Based on the estimates, the Baffin Bay sea ice area budget in terms of inflow and outflow are quantified and possible causes for its interannual variations and trends are analyzed. On average, the annual (September–August) inflows through the northern gate and Lancaster Sound are on the order of 205.8(±74.7)×103 km2 and 55.2(±17.8)×103 km2. In particular, a comparison with published results seems to suggest that about 75 %–85 % of the inflow through the northern gates is newly formed ice produced in the recurring North Water Polynya (NOW), in addition to the inflow via Nares Strait and Jones Sound. Meanwhile, the mean outflow via the southern gate approaches 394.3(±110.2)×103 km2. The distinct interannual variability for ice area flux through the northern gate and southern gate is partly explained by wind forcing associated with cross-gate sea level pressure difference, with correlations of 0.62 and 0.68, respectively. Also, significant increasing trends are found for the annual sea ice area flux through the three gates, amounting to 38.9×103, 82.2×103, and 7.5×103 km2 decade−1 for the northern gate, southern gate, and Lancaster Sound. These trends are chiefly related to the increasing ice motion, which is associated with thinner ice owing to the warmer climate (i.e., higher surface air temperature and shortened freezing period) and increased air and water drag coefficients over the past decades.

A Pareto-Optimal Multi-filter Architecture on FPGA for Image Processing Applications

Conventional two-dimensional filters for digital image processing are computationally intensive and consume significant hardware resources on field-programmable gate array (FPGA). Multiple filters implemented at a very low area and power budget are essential requirements of modern embedded systems for performing diversified image processing tasks. The previously reported multi-filter architectures are infeasible, since they do not provide pareto-optimal solution in terms of flexibility, hardware cost and versatility. This work presents a pareto-optimal multi-filter architecture that cater to these three major design concerns. The high degree of flexibility is induced in design by the deployment of an efficient customized 2,3 tree structure. This implements successive odd filter kernels in 44 different combinations with the maximum \(9 \times 9\) size, without hardware blocking. These implemented filters are structurally optimized and versatile for realizing diverse filter types at a very low cost. The complete design occupies only 2319 Slices on Xilinx Artix-7 FPGA and performs at 111 fps for full HD resolution with a very low power consumption of 197.5 mW. With these attributes, the proposed multi-filter architecture implements multiple filters of diverse types at very low area and power; this makes it well suited for image processing applications in real time. Finally, the proposed architecture is validated using Digilent Nexys4 development board.

A Fresh View on the Microarchitectural Design of FPGA-Based RISC CPUs in the IoT Era

The Internet-of-Things (IoT) revolution has shaped a new application domain where low-power RISC architectures constitute the standard computational backbone. The current de-facto design practice for such architectures is to extend the ISA and the corresponding microarchitecture with custom instructions to efficiently manage the complex tasks imposed by IoT applications, i.e., augmented reality, artificial intelligence and autonomous driving, within narrow energy and area budgets. However, the new IoT application domain also offers a unique opportunity to revisit and optimize the RISC microarchitectural design flow from a more communication- and memory-centric viewpoint. This manuscript critically explores and optimizes the design of a RISC CPU front-end for IoT delivering a two-fold objective: (i) provide an optimized CPU microarchitecture; and (ii) present a set of three design guidelines to steer the implementation of IoT CPUs. The exploration sits on a newly proposed Systems-on-Chip (SoC) and RISC CPU implementing the RISC-V/IMF ISA and accounting for area, timing, and performance design metrics. Such SoC offers a reference design to evaluate pros and cons of different microarchitectural solutions. A wide combination of microarchitectures considering different branch prediction schemes, cache design architectures and on-chip bus solutions have been evaluated. The entire exploration is focused on the FPGA-based implementation due to the renewed interest for this technology demonstrated by both the research community and companies. We note that ARM launched the DesignStart FPGA program to make available the Cortex-M microcontrollers on Xilinx FPGAs in the form of IP blocks.

Scaling and balancing methane fluxes in a heterogeneous tundra ecosystem of the Lena River Delta

Methane fluxes on an active flood plain situated in the Siberian Lena River Delta were studied applying the eddy covariance method. During the growing season, the observed fluxes exhibited a great deal of temporal variability, which was largely the result of the pronounced spatial variability of soil and vegetation characteristics within the footprint. Explaining this variability was based on three data-driven modelling approaches: the automatically operating algorithms stepwise regression as well as neural network, and a mechanistic model, which utilised exponential relationships between the methane flux and both flux drivers soil temperature and friction velocity. A substantial improvement in model performance was achieved by applying footprint information in the form of relative contributions of three vegetation classes to the flux signal. This aspect indicates that the vegetation served as an integrated proxy for flux drivers, whose characteristics permanently varied according to the shifting source area. The neural network performed best in explaining the variability of the observed methane fluxes. However, validating the models’ generalisability revealed that the mechanistic model provided the most predictive power suggesting that this model best captured the causality between the methane flux and its drivers. After integrating the gap-filled time series, all models yielded footprint budgets that were similar in magnitude. These budgets, however, lacked representativity due to the sensor location bias, i.e. their strong dependence on tower location, measurement height and wind field conditions. Thus, an unbiased budget of the total area of the flood plain was estimated utilising the mechanistic model. Initially, a downscaling procedure partitioned the observed flux with a seasonal mean of 0.012 μmol m-2 s-1 into three individual vegetation class fluxes accounting for shrubs (0.0004 μmol m-2 s-1), sedges (0.052 μmol m-2 s-1) and intermediate vegetation (0.018 μmol m-2 s-1). These decomposed fluxes in turn formed the basis – in conjunction with a classified high-resolution orthomosaic of the flood plain – for the vegetation class area-weighted upscaling. Alternatively, the straightforward upscaling of the footprint budgets (without the preceding downscaling) yielded budgets that underestimated the methane source strength of the flood plain by roughly 42 %. Hence, the application of fine-scale information on surface characteristics is crucial for both modelling methane flux dynamics and adequately estimating budgets of heterogeneous ecosystems being abundant in the tundra biome.

A Case of On-Chip Memory Subsystem Design for Low-Power CNN Accelerators

The rapid development of machine learning is enabling a plenty of novel applications, such as image and speech recognition for embedded and mobile devices. However, state-of-the-art deep learning models like convolutional neural networks (CNNs) are demanding so much on-chip storage and compute resources that they cannot be smoothly handled by low-power mobile or embedded systems. In order to fit large CNN models into mobile or more cutting-edge devices for IoT or cyberphysics applications, we proposed an efficient on-chip memory architecture for CNN inference acceleration, and showed its application to in-house single-instruction multiple-data structure machine learning processor. The redesigned on-chip memory subsystem, Memsqueezer, includes an active weight buffer and data buffer set that embraces specialized compression methods to reduce the footprint of CNN parameters (weights) and activation data, respectively. Memsqueezer buffer can compress the data and weight set according to the dataflow in computation, and it also includes a built-in redundancy detection mechanism that actively scans through the working-set of CNNs to boost their inference performance by eliminating the computation redundancy in CNN models. In our experiments, it is shown that the CNN processors with Memsqueezer buffers achieve more than $2{\times }$ performance improvement and reduces 85% energy consumption on average over the conventional buffer design with the same area budget.

Quantification by numerical simulation of the impact of gullies on the water budget of a basement area, southeastern Brazil

ABSTRACTMany regions of the world are affected by high density of large gullies that can impact on water resources. As these impacts have not been properly evaluated, this study presents a quantification model by numerical simulation of the effect of gullying on the water budget. An active gully of a very eroded region (Bação Complex) was selected and systematically monitored during a hydrological year. The simulation of the gully area by finite element method enabled the quantification of impacts, such as baseflow reduction and groundwater drawdown in the vicinity of this erosional feature. The baseflow reduction, when extrapolated to the whole complex, showed a significant baseflow loss during the dry seasons and an increase of stream flows during the rainy seasons, favouring floods. The numerical simulation results indicate that these modifications are relevant and must be considered when managing aquifers in intensely gullied areas with problems related to water availability.

Recovery of Holothuroidea population density, community composition, and respiration activity after a deep‐sea disturbance experiment

AbstractMining polymetallic nodules on abyssal plains will have adverse impacts on deep‐sea ecosystems, but it is largely unknown whether the impacted ecosystem will recover, and if so at what rate. In 1989 the “DISturbance and reCOLonization” (DISCOL) experiment was conducted in the Peru Basin where the seafloor was disturbed with a plough harrow construction to explore the effect of small‐scale sediment disturbance from deep‐sea mining. Densities of Holothuroidea in the region were last investigated 7 yr post‐disturbance, before 19 yr later, the DISCOL site was re‐visited in 2015. An “ocean floor observatory system” was used to photograph the seabed across ploughed and unploughed seafloor and at reference sites. The images were analyzed to determine the Holothuroidea population density and community composition, which were combined with in situ respiration measurements of individual Holothuroidea to generate a respiration budget of the study area. For the first time since the experimental disturbance, similar Holothuroidea densities were observed at the DISCOL site and at reference sites. The Holothuroidea assemblage was dominated by Amperima sp., Mesothuria sp., and Benthodytes typica, together contributing 46% to the Holothuroidea population density. Biomass and respiration were similar among sites, with a Holothuroidea community respiration of 5.84 10−4 ± 8.74 10−5 mmol C m−2 d−1 at reference sites. Although these results indicate recovery of Holothuroidea, extrapolations regarding recovery from deep‐sea mining activities must be made with caution: results presented here are based on a relatively small‐scale disturbance experiment as compared to industrial‐scale nodule mining, and also only represent one taxonomic class of the megafauna.

Quantification of climatic feedbacks on the Caspian Sea level variability and impacts from the Caspian Sea on the large-scale atmospheric circulation

With a fall of the Caspian Sea level (CSL), its size gets smaller and therefore the total evaporation over the sea is reduced. With a reduced evaporation from the sea, the fall of the CSL is weakened. This creates a negative feedback as less evaporation leads to less water losses of the Caspian Sea (CS). On the other hand, less evaporation reduces the water in the atmosphere, which may lead to less precipitation in the catchment area of the CS. The two opposite feedbacks are estimated by using an atmospheric climate model coupled with an ocean model only for the CS with different CS sizes while keeping all other forcings like oceanic sea surface temperatures (SSTs) and leaf area index the same from a global climate simulation. The investigation is concentrated on the medieval period because at that time the CSL changed dramatically from about − 30 to − 19 m below the mean ocean sea level, partly man-made. Models used for simulating the last millennium are not able to change the size of the CS dynamically so far. When results from such simulations are used to investigate the CSL variability and its causes, the present study should help to parameterize its feedbacks. A first assumption that the total evaporation from the CS will vary with the size of the CS (number of grid points representing the sea) is generally confirmed with the model simulations. The decrease of grid points from 15 to 14, 10, 8 or 7 leads to a decrease of evaporation to 96, 77, 70 and 54%. The lower decrease than initially expected from the number of grid points (93, 67, 53 and 47%) is probably due to the fact that there would also be some evaporation at grid points that run dry with a lower CSL but a cooling of the CS SST with increasing CS size in summer may be more important. The reduction of evaporation over the CS means more water for the budget of the whole catchment of the CS (an increase of the CSL) but from the gain through reduced evaporation over the CS, only 70% is found to remain in the water budget of the whole catchment area due to feedbacks with the precipitation. This suggests a high proportion of recycling of water within the CS catchment area. When using a model which does not have a correct CS size, the effect of a reduced CS area on the water budget for the whole CS catchment can be estimated by taking the evaporation over the sea multiplied by the proportional changed area. However, only 50% of that change is ending up in the water balance of the total catchment of the CS. A formula is provided. This method has been applied to estimate the CSL during the Last Glacial Maximum to be at − 30 to − 33 m. The experiments show as well that the CS has an impact on the large-scale atmospheric circulation with a widened Aleutian 500 hPa height field trough with increasing CS sizes. It is possible to validate this aspect with observational data.

Analysis of evapotranspiration components of a rainfed olive orchard during three contrasting years in a semi-arid climate

Evapotranspiration is one of the most important fluxes of the water budget in semi-arid areas. The estimation of actual crop transpiration is a major issue in those regions due to its remarkable impacts on the precision of irrigation scheduling, crop growth and yield. Rainfed olive trees are adapted to the southern part of the Mediterranean basin even though they are vulnerable to an increased number of drought spells that might occur under current climate change scenarios. This present paper studies both water and energy exchanges over a rainfed olive grove in semi-arid conditions. The hydrological functioning of sparse olive trees is difficult to characterize because of its low LAI. To better understand water exchanges within the Soil–Plant–Atmosphere continuum and better evaluate the evapotranspiration and its components, we combine data arising from eddy covariance, soil water content measurements and the sap flow method. First, we check the consistency of the evapotranspiration partitioning and water balance over three contrasted years: one wet and two dry. Total evapotranspiration (ET) from eddy covariance method compares well with the sum of the evaporation (E) generated from the surface soil moisture measurements and the transpiration derived from the sap flow method. The top meter soil water balance corresponds roughly to ET during the wet year but for the dry years there is an evidence of extraction by roots below the first meter of soil. Inter-annual variations of the transpiration and associated water stress levels are analyzed by the combined use of different types of eco-physiological (sap flow) as well as remotely sensed variables that can be monitored through proxi-detection (albedo, surface temperature, surface soil moisture). The amount and timing of vegetation stress are consistent throughout the various indicators. Consequently, this consistent set of data can be used to constrain a SVAT land-surface model capable of representing the various features of the water and energy budget for this specific land cover.

Majority Voting-Based Reduced Precision Redundancy Adders

Fault-tolerant systems are usually implemented with triple modular redundancy (TMR)-based protection techniques which has a huge area and power overhead. Previous works have shown that this overhead can be improved by using reduced precision redundancy (RPR)-based approaches for intrinsically precision-tolerant applications like image and video processing. The error detection and steering logic required in the RPR schemes utilize subtraction and comparisons which are more complex hardware operations than TMR voting. This overhead is affordable for complex hardware structures, for example, finite impulse response filters, but for simple structures like adders, it can be significant. In this letter, it is shown that combining the traditional TMR voting with the RPR brings up considerable savings in area and power budget when applied to very-large-scale integration adder circuits.

Sedimentary budget of the Southwest Sub‐basin, South China Sea: Controlling factors and geological implications

Calculation of the sediment budget in the South China Sea abyssal basin lacks constraints from drilling data and reinterpretation of seismic data. On the basis of six multichannel seismic profiles across the Southwest Sub‐basin (SWSB) and the drilling data from International Ocean Discovery Program Expedition 349, we divided the Cenozoic sediments into four sedimentary units. The sedimentary budget of the abyssal basin at different geological times was calculated. The previous works in the Mekong continental shelf and slope areas were integrated to calculate the sediment budget of the whole SWSB. Our researches show that the sediment budget in the SWSB increased during the Palaeogene and reached its first peak because of intensified erosion, which might have resulted from the continued uplift of the Tibetan Plateau and accelerated southeastward extrusion of the Indo‐China Peninsula. Since the Late Miocene, the sediment budget was mainly influenced by the East Asia monsoon, that is, the intensified winter monsoon decreased the sediment budget of the entire area during the Late Miocene, whereas the strengthened summer monsoon increased the sediment budget during the Pliocene and reached a peak during the Pleistocene. The sediment budgets of the Mekong continental shelf, the Mekong continental slope, and the abyssal basin have distinct characteristics, relating to the infilling sequence of the terrestrial sediments in different regions. The sediment provenance of the SWSB was mainly from the Indo‐China Peninsula, the Nansha area, and the Palawan before the Late Miocene. After that time, sediments were mainly transported from the modern Mekong River.

The Search for a Euro Area Safe Asset

This paper evaluates four approaches to creating or asset portfolios for the euro area: (1) a diversified portfolio of senior tranches of sovereign debt (national tranching); (2) a senior security backed by a diversified pool of national sovereign debt (ESBies); (3) debt issued by a senior financial intermediary, backed by a diversified pool of national debt (E-bonds); and (4) debt issued by a euro area budget or a leveraged wealth fund, based on member state contributions or dedicated direct revenue sources. None of these approaches envisages explicit guarantees by member states, and all could potentially produce safe assets in sufficient quantities to replace euro area sovereign bond holdings in euro area banks. At the same time, the four approaches differ across several important dimensions. A euro area budget or wealth fund could create the largest volume of safe assets, followed by ESBies, E-bonds, and national tranching. A euro area budget or wealth fund is also likely to have the lowest impact on the structure and liquidity of national bond markets, while national tranching would have the largest impact. ESBies and E-bonds occupy an intermediate position. ESBies and potentially bonds issued by a euro area budget would offer their holders greater protection from deep national defaults than the other two proposals. Both ESBies and national tranching would avoid cross-country redistribution by construction, whereas E-bonds and a euro area budget could have significant distributional consequences, depending on their design. E-bonds are unique in that they would raise the marginal cost of sovereign debt issuance at higher levels of debt, thereby exerting fiscal discipline, without necessarily raising average debt costs for lower-rated borrowers.

A Hybrid Instruction Prefetching Mechanism for Ultra Low-Power Multicore Clusters

The instruction memory hierarchy plays a critical role in performance and energy efficiency of ultralow-power (ULP) processors for the Internet-of-Things (IoT) end-nodes. This is mainly due to the extremely tight power envelope and area budgets, which imply small instruction-caches (I-Cache) operating at very low supply voltages (near-threshold). The challenge is aggravated by the fact that multiple processors, fetching in parallel, require plenty of bandwidth from the I-Caches. In this letter, we propose a low-cost and energy efficient hybrid instruction-prefetching mechanism to be integrated with a ULP multicore cluster. We study its performance for a wide range of IoT applications, from cryptography to computer vision, and show that it can effectively improve the hit-rate of almost all of them to above 95% (average performance improvement of over $2 \times $ ). In addition, we designed our prefetcher and integrated it in a 4-cores cluster in 28 nm fully-depleted silicon-on-insulator (FDSOI) technology. We show that system’s power consumption increases only by about 11% and silicon area by less than 1%. Altogether, a total energy reduction of $1.9\times $ is achieved, thanks to more than $2 \times $ performance improvement, enabling a significantly longer battery life.

A DFA-Resistant and Masked PRESENT with Area Optimization for RFID Applications

Radio-Frequency Identification (RFID) tag-based applications are usually resource constrained and security sensitive. However, only about 2,000 gate equivalents in a tag can be budgeted for implementing security components [27]. This requires not only lightweight cryptographic algorithms such as PRESENT (around 1,000 gate equivalents) but also lightweight protections against modern Side Channel Attacks (SCAs). With this budget, the first-order masking and fault detection are two suitable countermeasures to be developed for PRESENT. However, if both countermeasures are applied without any optimization, it will significantly exceed the given area budget. In this work, we optimize area to include both countermeasures to maximize the security for PRESENT within this RFID area budget. The most area-consuming parts of the proposed design are the masked S-boxes and the inverse masked S-boxes. To optimize the area, we have deduced a computational relationship between these two parts, which enables us to reuse the hardware resource of the masked S-boxes to implement the inverse masked S-boxes. The proposed design takes up only 2,376 gates with UMC 65nm CMOS technology. Compared with the unoptimized design, our implementation reduces the overall area by 28.45%. We have tested the effectiveness of the first-order Differential Power Analysis (DPA) and Differential Fault Analysis (DFA) -resistant countermeasures. Experimental results show that we have enhanced the SCA resistance of our PRESENT implementation.

Development of electrical conductivity measurement technology for key plant physiological information using microneedle sensor

Impedance measurement is a widely used technique for monitoring ion species in various applications. In plant cultivation, the impedance system is used to measure the electrical conductivity (EC) of nutrient solutions. Recent research has shown that the quality and quantity of horticultural crops, e.g. tomato, can be optimized by controlling the salinity of nutrient solutions. However, understanding the detailed response of a plant to a nutrient solution is not possible until the fruit is fully grown or by sacrificing the stem. To overcome this issue, horticultural crop cultivation requires real-time monitoring of the EC inside the stem. Using this data, the growth model of a plant could be constructed, and the response of the plant to external environment determined. In this paper, we propose an implantable microneedle device equipped with a micro-patterned impedance measurement system for direct measurement of the EC inside the tomato stem. The fabrication process includes silicon-based steps such as microscale deposition, photolithography, and a deep etching process. Further, microscale fabrication enables all functional elements to fulfill the area budget and be very accurate with minimal plant invasion. A two-electrode geometry is used to match the measurement condition of the tomato stem. Real-time measurement of local sap condition inside the plant in which real-time data for tomato sap EC is obtained after calibration at various concentrations of standard solution demonstrate the efficacy of the proposed device.

An Integrated Vector-Scalar Design on an In-Order ARM Core

In the low-end mobile processor market, power, energy, and area budgets are significantly lower than in the server/desktop/laptop/high-end mobile markets. It has been shown that vector processors are a highly energy-efficient way to increase performance; however, adding support for them incurs area and power overheads that would not be acceptable for low-end mobile processors. In this work, we propose an integrated vector-scalar design for the ARM architecture that mostly reuses scalar hardware to support the execution of vector instructions. The key element of the design is our proposed block-based model of execution that groups vector computational instructions together to execute them in a coordinated manner. We implemented a classic vector unit and compare its results against our integrated design. Our integrated design improves the performance (more than 6×) and energy consumption (up to 5×) of a scalar in-order core with negligible area overhead (only 4.7% when using a vector register with 32 elements). In contrast, the area overhead of the classic vector unit can be significant (around 44%) if a dedicated vector floating-point unit is incorporated. Our block-based vector execution outperforms the classic vector unit for all kernels with floating-point data and also consumes less energy. We also complement the integrated design with three energy/performance-efficient techniques that further reduce power and increase performance. The first proposal covers the design and implementation of chaining logic that is optimized to work with the cache hierarchy through vector memory instructions, the second proposal reduces the number of reads/writes from/to the vector register file, and the third idea optimizes complex memory access patterns with the memory shape instruction and unified indexed vector load.

SPARCNet

Deep neural networks have been shown to outperform prior state-of-the-art solutions that often relied heavily on hand-engineered feature extraction techniques coupled with simple classification algorithms. In particular, deep convolutional neural networks have been shown to dominate on several popular public benchmarks such as the ImageNet database. Unfortunately, the benefits of deep networks have yet to be fully exploited in embedded, resource-bound settings that have strict power and area budgets. Graphical processing unit (GPU) have been shown to improve throughput and energy-efficiency over central processing unit (CPU) due to their highly parallel architecture yet still impose a significant power burden. In a similar fashion, field programmable gate array (FPGA) can be used to improve performance while further allowing more fine-grained control over implementation to improve efficiency. In order to reduce power and area while still achieving required throughput, classification-efficient network architectures are required in addition to optimal deployment on efficient hardware. In this work, we target both of these enterprises. For the first objective, we analyze simple, biologically inspired reduction strategies that are applied both before and after training. The central theme of the techniques is the introduction of sparsification to help dissolve away the dense connectivity that is often found at different levels in convolutional neural networks. The sparsification techniques include feature compression partition , structured filter pruning , and dynamic feature pruning . Additionally, we explore filter factorization and filter quantization approximation techniques to further reduce the complexity of convolutional layers. In the second contribution, we propose SPARCNet, a hardware accelerator for efficient deployment of SPAR se C onvolutional NET works. The accelerator looks to enable deploying networks in such resource-bound settings by both exploiting efficient forms of parallelism inherent in convolutional layers and by exploiting the sparsification and approximation techniques proposed. To demonstrate both contributions, modern deep convolutional network architectures containing millions of parameters are explored within the context of the computer vision dataset CIFAR. Utilizing the reduction techniques, we demonstrate the ability to reduce computation and memory by 60% and 93% with less than 0.03% impact on accuracy when compared to the best baseline network with 93.47% accuracy. The SPARCNet accelerator with different numbers of processing engines is implemented on a low-power Artix-7 FPGA platform. Additionally, the same networks are optimally implemented on a number of embedded commercial-off-the-shelf platforms including NVIDIAs CPU+GPU SoCs TK1 and TX1 and Intel Edison. Compared to NVIDIAs TK1 and TX1, the FPGA-based accelerator obtains 11.8 × and 7.5 × improvement in energy efficiency while maintaining a classification throughput of 72 images/s. When further compared to a number of recent FPGA-based accelerators, SPARCNet is able to achieve up to 15 × improvement in energy efficiency while consuming less than 2W of total board power at 100MHz. In addition to improving efficiency, the accelerator has built-in support for sparsification techniques and ability to perform in-place rectified linear unit (ReLU) activation function, max-pooling, and batch normalization.