Showing 1–17 of 17 results for author: Hassoun, J

PROV-IO+: A Cross-Platform Provenance Framework for Scientific Data on HPC Systems

Authors: Runzhou Han, Mai Zheng, Suren Byna, Houjun Tang, Bin Dong, Dong Dai, Yong Chen, Dongkyun Kim, Joseph Hassoun, David Thorsley, Matthew Wolf

Abstract: Data provenance, or data lineage, describes the life cycle of data. In scientific workflows on HPC systems, scientists often seek diverse provenance (e.g., origins of data products, usage patterns of datasets). Unfortunately, existing provenance solutions cannot address the challenges due to their incompatible provenance models and/or system implementations. In this paper, we analyze four represen… ▽ More Data provenance, or data lineage, describes the life cycle of data. In scientific workflows on HPC systems, scientists often seek diverse provenance (e.g., origins of data products, usage patterns of datasets). Unfortunately, existing provenance solutions cannot address the challenges due to their incompatible provenance models and/or system implementations. In this paper, we analyze four representative scientific workflows in collaboration with the domain scientists to identify concrete provenance needs. Based on the first-hand analysis, we propose a provenance framework called PROV-IO+, which includes an I/O-centric provenance model for describing scientific data and the associated I/O operations and environments precisely. Moreover, we build a prototype of PROV-IO+ to enable end-to-end provenance support on real HPC systems with little manual effort. The PROV-IO+ framework can support both containerized and non-containerized workflows on different HPC platforms with flexibility in selecting various classes of provenance. Our experiments with realistic workflows show that PROV-IO+ can address the provenance needs of the domain scientists effectively with reasonable performance (e.g., less than 3.5% tracking overhead for most experiments). Moreover, PROV-IO+ outperforms a state-of-the-art system (i.e., ProvLake) in our experiments. △ Less

Submitted 1 August, 2023; originally announced August 2023.

SaiT: Sparse Vision Transformers through Adaptive Token Pruning

Authors: Ling Li, David Thorsley, Joseph Hassoun

Abstract: While vision transformers have achieved impressive results, effectively and efficiently accelerating these models can further boost performances. In this work, we propose a dense/sparse training framework to obtain a unified model, enabling weight sharing across various token densities. Thus one model offers a range of accuracy and throughput tradeoffs for different applications. Besides, we intro… ▽ More While vision transformers have achieved impressive results, effectively and efficiently accelerating these models can further boost performances. In this work, we propose a dense/sparse training framework to obtain a unified model, enabling weight sharing across various token densities. Thus one model offers a range of accuracy and throughput tradeoffs for different applications. Besides, we introduce adaptive token pruning to optimize the patch token sparsity based on the input image. In addition, we investigate knowledge distillation to enhance token selection capability in early transformer modules. Sparse adaptive image Transformer (SaiT) offers varying levels of model acceleration by merely changing the token sparsity on the fly. Specifically, SaiT reduces the computation complexity (FLOPs) by 39% - 43% and increases the throughput by 67% - 91% with less than 0.5% accuracy loss for various vision transformer models. Meanwhile, the same model also provides the zero accuracy drop option by skipping the sparsification step. SaiT achieves better accuracy and computation tradeoffs than state-of-the-art transformer and convolutional models. △ Less

Submitted 11 October, 2022; originally announced October 2022.

MaiT: Leverage Attention Masks for More Efficient Image Transformers

Authors: Ling Li, Ali Shafiee Ardestani, Joseph Hassoun

Abstract: Though image transformers have shown competitive results with convolutional neural networks in computer vision tasks, lacking inductive biases such as locality still poses problems in terms of model efficiency especially for embedded applications. In this work, we address this issue by introducing attention masks to incorporate spatial locality into self-attention heads. Local dependencies are cap… ▽ More Though image transformers have shown competitive results with convolutional neural networks in computer vision tasks, lacking inductive biases such as locality still poses problems in terms of model efficiency especially for embedded applications. In this work, we address this issue by introducing attention masks to incorporate spatial locality into self-attention heads. Local dependencies are captured efficiently with masked attention heads along with global dependencies captured by unmasked attention heads. With Masked attention image Transformer - MaiT, top-1 accuracy increases by up to 1.7% compared to CaiT with fewer parameters and FLOPs, and the throughput improves by up to 1.5X compared to Swin. Encoding locality with attention masks is model agnostic, and thus it applies to monolithic, hierarchical, or other novel transformer architectures. △ Less

Submitted 6 July, 2022; originally announced July 2022.

Scalable Composites Benefiting from Transition-Metal Oxides as Cathode Materials for Efficient Lithium-Sulfur Batteries

Authors: Vittorio Marangon, Eugenio Scaduti, Viviana Fatima Vinci, Jusef Hassoun

Abstract: Composite materials achieved by including transition-metal oxides with different structures and morphologies in sulfur are suggested as scalable cathodes for high-energy lithium-sulfur (Li-S) batteries. The composites contain 80 wt.% sulfur and 20 wt.% of either MnO2 or TiO2, leading to a sulfur content in the electrode of 64 wt.% and revealing a reversible, fast, and lowly polarized conversion pr… ▽ More Composite materials achieved by including transition-metal oxides with different structures and morphologies in sulfur are suggested as scalable cathodes for high-energy lithium-sulfur (Li-S) batteries. The composites contain 80 wt.% sulfur and 20 wt.% of either MnO2 or TiO2, leading to a sulfur content in the electrode of 64 wt.% and revealing a reversible, fast, and lowly polarized conversion process in the cell with limited interphase resistance. The S-TiO2 composite exhibits an excellent rate capability between C/10 and 2C, and a cycle life extended over 400 cycles at 2C, owing to the effects of the nanometric TiO2 additive in boosting the reaction kinetics. Instead, the micrometric sized particles of MnO2 partially limit the electrochemical activity of S-MnO2 to the current rate of 1C. Nevertheless, both S-MnO2 and S-TiO2 withstand a sulfur loading up to values approaching 6 mgcm-2, and deliver an areal capacit ranging from about 4.5 to 5.5 mAhcm-2 at C/5. The excellent performances of the metal oxide-sulfur electrodes, even at high active material loading, and the possible scalability of the synthetic pathway adopted in the work suggest that the composites are viable cathodes for next-generation Li-S batteries with high energy density and efficient electrochemical process. △ Less

Submitted 16 June, 2022; originally announced June 2022.

Journal ref: ChemElectroChem 2022, 9, e202200374

A Fast Post-Training Pruning Framework for Transformers

Authors: Woosuk Kwon, Sehoon Kim, Michael W. Mahoney, Joseph Hassoun, Kurt Keutzer, Amir Gholami

Abstract: Pruning is an effective way to reduce the huge inference cost of Transformer models. However, prior work on pruning Transformers requires retraining the models. This can add high training cost and high complexity to model deployment, making it difficult to use in many practical situations. To address this, we propose a fast post-training pruning framework for Transformers that does not require any… ▽ More Pruning is an effective way to reduce the huge inference cost of Transformer models. However, prior work on pruning Transformers requires retraining the models. This can add high training cost and high complexity to model deployment, making it difficult to use in many practical situations. To address this, we propose a fast post-training pruning framework for Transformers that does not require any retraining. Given a resource constraint and a sample dataset, our framework automatically prunes the Transformer model using structured sparsity methods. To retain high accuracy without retraining, we introduce three novel techniques: (i) a lightweight mask search algorithm that finds which heads and filters to prune based on the Fisher information; (ii) mask rearrangement that complements the search algorithm; and (iii) mask tuning that reconstructs the output activations for each layer. We apply our method to BERT-base and DistilBERT, and we evaluate its effectiveness on GLUE and SQuAD benchmarks. Our framework achieves up to 2.0x reduction in FLOPs and 1.56x speedup in inference latency, while maintaining < 1% loss in accuracy. Importantly, our framework prunes Transformers in less than 3 minutes on a single GPU, which is over two orders of magnitude faster than existing pruning approaches that retrain the models. △ Less

Submitted 17 October, 2022; v1 submitted 29 March, 2022; originally announced April 2022.

Comments: NeurIPS 2022

Synthesis and Characterization of a LiFe0.6Mn0.4PO4 Olivine Cathode for Application in a New Lithium Polymer Battery

Authors: Luca Minnetti, Vittorio Marangon, Jusef Hassoun

Abstract: A LiFe0.6Mn0.4PO4 (LFMP) cathode exploiting the olivine structure is herein synthesized and characterized in terms of structure, morphology, and electrochemical features in a lithium cell. The material shows reversibility of the electrochemical process which evolves at 3.5 and 4 V versus Li+/Li due to the Fe+2/Fe+3 and Mn+2/Mn+3 redox couples, respectively, as determined by cyclic voltammetry. The… ▽ More A LiFe0.6Mn0.4PO4 (LFMP) cathode exploiting the olivine structure is herein synthesized and characterized in terms of structure, morphology, and electrochemical features in a lithium cell. The material shows reversibility of the electrochemical process which evolves at 3.5 and 4 V versus Li+/Li due to the Fe+2/Fe+3 and Mn+2/Mn+3 redox couples, respectively, as determined by cyclic voltammetry. The LFMP has a well-defined olivine structure revealed by X-ray diffraction, a morphology consisting of submicron particle aggregated into micrometric clusters as indicated by scanning and transmission electron microscopy, with a carbon weight ratio of about 4% as suggested by thermogravimetry. The electrode is used in lithium cells subjected to galvanostatic cycling with a conventional liquid electrolyte, and demonstrates a maximum capacity of 130 mAh g-1, satisfactory rate capability, excellent efficiency, and a stable trend. Therefore, the material is studied in a lithium metal polymer cell exploiting an electrolyte based on polyethylene glycol dimethyl ether with a solid configuration. The cell reveals very promising features in terms of capacity, efficiency, and retention, and suggests the LFMP material as a suitable electrode for polymer batteries characterized by increased energy density and remarkable safety. △ Less

Submitted 10 March, 2022; originally announced March 2022.

Journal ref: Adv. Sustainable Syst. 2022, 6, 2100464

Glyme-based electrolytes: suitable solutions for next-generation lithium batteries

Authors: Daniele Di Lecce, Vittorio Marangon, Hun-Gi Jung, Yoichi Tominaga, Steve Greenbaum, Jusef Hassoun

Abstract: The concept of green in a battery involves the chemical nature of electrodes and electrolytes as well as the economic sustainability of the cell. Although these aspects are typically discussed separately, they are deeply interconnected: indeed, a new electrolyte can allow the use of different cathodes with higher energy, lower cost or more pronounced environmental compatibility. We focus on altern… ▽ More The concept of green in a battery involves the chemical nature of electrodes and electrolytes as well as the economic sustainability of the cell. Although these aspects are typically discussed separately, they are deeply interconnected: indeed, a new electrolyte can allow the use of different cathodes with higher energy, lower cost or more pronounced environmental compatibility. We focus on alternative class of electrolyte solutions for lithium batteries formed by dissolving LiX salts in glyme solvents, i.e., organic ethers with the molecular formula CH3O[CH2CH2O]nCH3 differing by chain length. The advantages of these electrolytes are illustrated in terms of flammability, stability, toxicity, environmental compatibility, cell performances and economic impact. A particular light is shed on the stability of these systems, particularly in the polymer state, and in various environments including oxygen, sulfur and high-energy lithium metal. The most relevant studies on the chemical-physical features, the characteristic structures, the favorable properties, and the electrochemical behavior of the glyme-based solutions are discussed, and the most recent technological achievements in terms of cell design and battery performance are described. The use of glyme-based electrolytes in high-energy cells arranged by coupling the lithium-metal anode with conventional insertion cathodes as well as in alternative and new batteries exploiting the Li-S and Li-O2 conversion processes are described in detail. The paragraphs reveal bonuses, including safety, low cost and sustainability, that can be achieved by employing the glyme-based electrolytes with respect to the commercially available ones, in particular taking into account future and alternative applications. Particular relevance is given by the glymes with long chain that reveal a remarkable stability, high safety and very low toxicity. △ Less

Submitted 9 February, 2022; originally announced February 2022.

Novel Lithium-Sulfur Polymer Battery Operating at Moderate Temperature

Authors: Vittorio Marangon, Daniele Di Lecce, Luca Minnetti, Jusef Hassoun

Abstract: A safe lithium-sulfur (Li-S) battery employs a composite polymer electrolyte based on a poly(ethylene glycol) dimethyl ether (PEGDME) solid at room temperature. The electrolyte membrane enables a stable and reversible Li-S electrochemical process already at 50°C, with low resistance at the electrode/electrolyte interphase and fast Li+ transport. The relatively low molecular weight of the PEGDME an… ▽ More A safe lithium-sulfur (Li-S) battery employs a composite polymer electrolyte based on a poly(ethylene glycol) dimethyl ether (PEGDME) solid at room temperature. The electrolyte membrane enables a stable and reversible Li-S electrochemical process already at 50°C, with low resistance at the electrode/electrolyte interphase and fast Li+ transport. The relatively low molecular weight of the PEGDME and the optimal membrane composition in terms of salts and ceramic allow a liquid-like Li-S conversion reaction by heating at moderately high temperature, still holding the solid-like polymer state of the cell. Therefore, the electrochemical reaction of the polymer Li-S cell is characterized by the typical dissolution of lithium polysulfides into the electrolyte medium during discharge and the subsequent deposition of sulfur at the electrode/electrolyte interphase during charge. On the other hand, the remarkable thermal stability of the composite polymer electrolyte (up to 300°C) suggests a lithium-metal battery with safety content significantly higher than that using the common, flammable liquid solutions. Hence, the Li-S polymer battery delivers at 50°C and 2 V a stable capacity approaching 700 mAhgS-1, with a steady-state coulombic efficiency of 98%. These results suggest a novel, alternative approach to achieve safe, high energy batteries with solid polymer configuration. △ Less

Submitted 26 October, 2021; originally announced October 2021.

Journal ref: ChemElectroChem 8 (2021) 3971-3981

Degradation of Layered Oxide Cathode in a Sodium Battery: A Detailed Investigation by X-Ray Tomography at the Nanoscale

Authors: Daniele Di Lecce, Vittorio Marangon, Mark Isaacs, Robert Palgrave, Paul R. Shearing, Jusef Hassoun

Abstract: The degradation mechanism in a sodium cell of a layered Na0.48Al0.03Co0.18Ni0.18Mn0.47O2 (NCAM) cathode with P3/P2 structure is investigated by revealing the changes in microstructure and composition upon cycling. The work aims to rationalize the gradual performance decay and the alteration of the electrochemical response in terms of polarization, voltage signature, and capacity loss. Spatial reco… ▽ More The degradation mechanism in a sodium cell of a layered Na0.48Al0.03Co0.18Ni0.18Mn0.47O2 (NCAM) cathode with P3/P2 structure is investigated by revealing the changes in microstructure and composition upon cycling. The work aims to rationalize the gradual performance decay and the alteration of the electrochemical response in terms of polarization, voltage signature, and capacity loss. Spatial reconstructions of the electrode by X-ray computed tomography at the nanoscale supported by quantitative and qualitative analyses show fractures and deformations in the cycled layered metal-oxide particles, as well as inorganic side compounds deposited on the material. These irreversible morphological modifications reflect structural heterogeneities across the cathode particles due to formation of various domains with different Na+ intercalation degrees. Besides, X-ray photoelectron spectroscopy data suggest that the latter inorganic species in the cycled electrode are mainly composed of NaF, Na2O, and NaCO3 formed by parasitic electrolyte decomposition. The precipitation of these insulating compounds at the electrode/electrolyte interphase and the related structural stresses induced in the material lead to a decrease in cathode particle size and partial loss of electrochemical activity. The retention of the NCAM phase after cycling suggests that electrolyte upgrade may improve the performance of the cathode to achieve practical application for sustainable energy storage. △ Less

Submitted 25 August, 2021; originally announced August 2021.

Journal ref: Small Methods 5 (2021) 2100596

Synthesis of a High-Capacity α-Fe2O3@C Conversion Anode and a High-Voltage LiNi0.5Mn1.5O4 Spinel Cathode and Their Combination in a Li-Ion Battery

Authors: Shuangying Wei, Daniele Di Lecce, Riccardo Messini D'Agostini, Jusef Hassoun

Abstract: A Li-conversion alpha-Fe2O3@C nanocomposite anode and a high-voltage LiNi0.5Mn1.5O4 cathode are synthesized in parallel, characterized, and combined in a Li-ion battery. alpha-Fe2O3@C is prepared via annealing of maghemite iron oxide and sucrose under an argon atmosphere and subsequent oxidation in air. The nanocomposite exhibits a satisfactory electrochemical response in a lithium half-cell, deli… ▽ More A Li-conversion alpha-Fe2O3@C nanocomposite anode and a high-voltage LiNi0.5Mn1.5O4 cathode are synthesized in parallel, characterized, and combined in a Li-ion battery. alpha-Fe2O3@C is prepared via annealing of maghemite iron oxide and sucrose under an argon atmosphere and subsequent oxidation in air. The nanocomposite exhibits a satisfactory electrochemical response in a lithium half-cell, delivering almost 900 mA h g-1, as well as a significantly longer cycle life and higher rate capability compared to the bare iron oxide precursor. The LiNi0.5Mn1.5O4 cathode, achieved using a modified co-precipitation approach, reveals a well-defined spinel structure without impurities, a sub-micrometrical morphology, and a reversible capacity of ca. 120 mA h g-1 in a lithium half-cell with an operating voltage of 4.8 V. Hence, a lithium-ion battery is assembled by coupling the alpha-Fe2O3@C anode with the LiNi0.5Mn1.5O4 cathode. This cell operates at about 3.2 V, delivering a stable capacity of 110 mA h g-1 (referred to the cathode mass) with a Coulombic efficiency exceeding 97%. Therefore, this cell is suggested as a promising energy storage system with expected low economic and environmental impacts. △ Less

Submitted 25 August, 2021; originally announced August 2021.

Journal ref: ACS Appl. Energy Mater. 4 (2021) 8340-8349

A Stable High-Capacity Lithium-Ion Battery Using a Biomass-Derived Sulfur-Carbon Cathode and Lithiated Silicon Anode

Authors: Vittorio Marangon, Celia Hernández-Renter, Mara Olivares-Marín, Vicente Gómez-Serrano, Álvaro Caballero, Julián Morales, Jusef Hassoun

Abstract: A full lithium-ion-sulfur cell with a remarkable cycle life was achieved by combining an environmentally sustainable biomass-derived sulfur-carbon cathode and a pre-lithiated silicon oxide anode. X-ray diffraction, Raman spectroscopy, energy dispersive spectroscopy, and thermogravimetry of the cathode evidenced the disordered nature of the carbon matrix in which sulfur was uniformly distributed wi… ▽ More A full lithium-ion-sulfur cell with a remarkable cycle life was achieved by combining an environmentally sustainable biomass-derived sulfur-carbon cathode and a pre-lithiated silicon oxide anode. X-ray diffraction, Raman spectroscopy, energy dispersive spectroscopy, and thermogravimetry of the cathode evidenced the disordered nature of the carbon matrix in which sulfur was uniformly distributed with a weight content as high as 75%, while scanning and transmission electron microscopy revealed the micrometric morphology of the composite. The sulfur-carbon electrode in the lithium half-cell exhibited a maximum capacity higher than 1200 mAhgS-1, reversible electrochemical process, limited electrode/electrolyte interphase resistance, and a rate capability up to C/2. The material showed a capacity decay of about 40% with respect to the steady-state value over 100 cycles, likely due to the reaction with the lithium metal of dissolved polysulfides or impurities including P detected in the carbon precursor. Therefore, the replacement of the lithium metal with a less challenging anode was suggested, and the sulfur-carbon composite was subsequently investigated in the full lithium-ion-sulfur battery employing a Li-alloying silicon oxide anode. The full-cell revealed an initial capacity as high as 1200 mAhgS-1, a retention increased to more than 79% for 100 galvanostatic cycles, and 56% over 500 cycles. The data reported herein well indicated the reliability of energy storage devices with extended cycle life employing high-energy, green, and safe electrode materials. △ Less

Submitted 25 August, 2021; originally announced August 2021.

Journal ref: ChemSusChem 14 (2021) 3333-3343

Griffin: Rethinking Sparse Optimization for Deep Learning Architectures

Authors: Jong Hoon Shin, Ali Shafiee, Ardavan Pedram, Hamzah Abdel-Aziz, Ling Li, Joseph Hassoun

Abstract: This paper examines the design space trade-offs of DNNs accelerators aiming to achieve competitive performance and efficiency metrics for all four combinations of dense or sparse activation/weight tensors. To do so, we systematically examine the overheads of supporting sparsity on top of an optimized dense core. These overheads are modeled based on parameters that indicate how a multiplier can bor… ▽ More This paper examines the design space trade-offs of DNNs accelerators aiming to achieve competitive performance and efficiency metrics for all four combinations of dense or sparse activation/weight tensors. To do so, we systematically examine the overheads of supporting sparsity on top of an optimized dense core. These overheads are modeled based on parameters that indicate how a multiplier can borrow a nonzero operation from the neighboring multipliers or future cycles. As a result of this exploration, we identify a few promising designs that perform better than prior work. Our findings suggest that even the best design targeting dual sparsity yields a 20%-30% drop in power efficiency when performing on single sparse models, i.e., those with only sparse weight or sparse activation tensors. We found that one can reuse resources of the same core to maintain high performance and efficiency when running single sparsity or dense models. We call this hybrid architecture Griffin. Griffin is 1.2, 3.0, 3.1, and 1.4X more power-efficient than state-of-the-art sparse architectures, for dense, weight-only sparse, activation-only sparse, and dual sparse models, respectively. △ Less

Submitted 1 November, 2021; v1 submitted 27 July, 2021; originally announced July 2021.

Comments: Accepted at the 28th IEEE International Symposium on High-Performance Computer Architecture - HPCA 2022

Lithium-Metal Batteries Using Sustainable Electrolyte Media and Various Cathode Chemistries

Authors: Vittorio Marangon, Luca Minnetti, Matteo Adami, Alberto Barlini, Jusef Hassoun

Abstract: Lithium-metal batteries employing concentrated glyme-based electrolytes and different cathode chemistries are herein evaluated in view of a safe use of the highly energetic alkali-metal anode. Indeed, diethylene-glycol dimethyl-ether (DEGDME) and triethylene-glycol dimethyl-ether (TREGDME) dissolving lithium bis(trifluoromethanesulfonyl)imide (LiTFSI) and lithium nitrate (LiNO3) in concentration a… ▽ More Lithium-metal batteries employing concentrated glyme-based electrolytes and different cathode chemistries are herein evaluated in view of a safe use of the highly energetic alkali-metal anode. Indeed, diethylene-glycol dimethyl-ether (DEGDME) and triethylene-glycol dimethyl-ether (TREGDME) dissolving lithium bis(trifluoromethanesulfonyl)imide (LiTFSI) and lithium nitrate (LiNO3) in concentration approaching the solvents saturation limit are used in lithium batteries employing either a conversion sulfur-tin composite (S:Sn 80:20 w/w) or a Li+ (de-)insertion LiFePO4 cathode. Cyclic voltammetry (CV) and electrochemical impedance spectroscopy (EIS) clearly show the suitability of the concentrated electrolytes in terms of process reversibility and low interphase resistance, particularly upon a favorable activation. Galvanostatic measurements performed in the lithium-sulfur (Li/S) batteries reveal promising capacities at room temperature (25 °C) and a value as high as 1300 mAh gS-1 for DEGDME-based electrolyte at 35 °C. On the other hand, the lithium-LiFePO4 (Li/LFP) cells exhibit satisfactory cycling behavior, in particular when employing an additional reduction step at low voltage cutoff (i.e., 1.2 V) during the first discharge to consolidate the solid electrolyte interphase (SEI). This procedure allows a coulombic efficiency near 100 %, a capacity approaching 160 mAh g-1 and relevant retention particularly for the cell using TREGDME-based electrolyte. Therefore, this work suggests the use of concentrated glyme-based electrolytes, the fine tuning of the operative conditions, and the careful selection of active materials chemistry as significant steps to achieve practical and safe lithium-metal batteries. △ Less

Submitted 9 July, 2021; originally announced July 2021.

Journal ref: Energy Fuels 35 (2021) 10284-10292

Learned Token Pruning for Transformers

Authors: Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, Kurt Keutzer

Abstract: Deploying transformer models in practice is challenging due to their inference cost, which scales quadratically with input sequence length. To address this, we present a novel Learned Token Pruning (LTP) method which adaptively removes unimportant tokens as an input sequence passes through transformer layers. In particular, LTP prunes tokens with an attention score below a threshold value which is… ▽ More Deploying transformer models in practice is challenging due to their inference cost, which scales quadratically with input sequence length. To address this, we present a novel Learned Token Pruning (LTP) method which adaptively removes unimportant tokens as an input sequence passes through transformer layers. In particular, LTP prunes tokens with an attention score below a threshold value which is learned for each layer during training. Our threshold-based method allows the length of the pruned sequence to vary adaptively based on the input sequence, and avoids algorithmically expensive operations such as top-k token selection. We extensively test the performance of LTP on GLUE tasks and show that our method outperforms the prior state-of-the-art token pruning methods by up to ~2.5% higher accuracy with the same amount of FLOPs. In particular, LTP achieves up to 2.1x FLOPs reduction with less than 1% accuracy drop, which results in up to 1.9x and 2.0x throughput improvement on Intel Haswell CPUs and NVIDIA V100 GPUs, respectively. Furthermore, we demonstrate that LTP is more robust than prior methods to variations on input sentence lengths. Our code has been developed in PyTorch and has been open-sourced. △ Less

Submitted 2 June, 2022; v1 submitted 2 July, 2021; originally announced July 2021.

Comments: KDD 2022 (Research Track)

Rethinking Floating Point Overheads for Mixed Precision DNN Accelerators

Authors: Hamzah Abdel-Aziz, Ali Shafiee, Jong Hoon Shin, Ardavan Pedram, Joseph H. Hassoun

Abstract: In this paper, we propose a mixed-precision convolution unit architecture which supports different integer and floating point (FP) precisions. The proposed architecture is based on low-bit inner product units and realizes higher precision based on temporal decomposition. We illustrate how to integrate FP computations on integer-based architecture and evaluate overheads incurred by FP arithmetic su… ▽ More In this paper, we propose a mixed-precision convolution unit architecture which supports different integer and floating point (FP) precisions. The proposed architecture is based on low-bit inner product units and realizes higher precision based on temporal decomposition. We illustrate how to integrate FP computations on integer-based architecture and evaluate overheads incurred by FP arithmetic support. We argue that alignment and addition overhead for FP inner product can be significant since the maximum exponent difference could be up to 58 bits, which results into a large alignment logic. To address this issue, we illustrate empirically that no more than 26-bitproduct bits are required and up to 8-bit of alignment is sufficient in most inference cases. We present novel optimizations based on the above observations to reduce the FP arithmetic hardware overheads. Our empirical results, based on simulation and hardware implementation, show significant reduction in FP16 overhead. Over typical mixed precision implementation, the proposed architecture achieves area improvements of up to 25% in TFLOPS/mm2and up to 46% in TOPS/mm2with power efficiency improvements of up to 40% in TFLOPS/Wand up to 63% in TOPS/W. △ Less

Submitted 27 January, 2021; originally announced January 2021.

Comments: Accepted to appear in 4th Conference on Machine Learning and Systems 2021

Post-Training Piecewise Linear Quantization for Deep Neural Networks

Authors: Jun Fang, Ali Shafiee, Hamzah Abdel-Aziz, David Thorsley, Georgios Georgiadis, Joseph Hassoun

Abstract: Quantization plays an important role in the energy-efficient deployment of deep neural networks on resource-limited devices. Post-training quantization is highly desirable since it does not require retraining or access to the full training dataset. The well-established uniform scheme for post-training quantization achieves satisfactory results by converting neural networks from full-precision to 8… ▽ More Quantization plays an important role in the energy-efficient deployment of deep neural networks on resource-limited devices. Post-training quantization is highly desirable since it does not require retraining or access to the full training dataset. The well-established uniform scheme for post-training quantization achieves satisfactory results by converting neural networks from full-precision to 8-bit fixed-point integers. However, it suffers from significant performance degradation when quantizing to lower bit-widths. In this paper, we propose a piecewise linear quantization (PWLQ) scheme to enable accurate approximation for tensor values that have bell-shaped distributions with long tails. Our approach breaks the entire quantization range into non-overlapping regions for each tensor, with each region being assigned an equal number of quantization levels. Optimal breakpoints that divide the entire range are found by minimizing the quantization error. Compared to state-of-the-art post-training quantization methods, experimental results show that our proposed method achieves superior performance on image classification, semantic segmentation, and object detection with minor overhead. △ Less

Submitted 18 March, 2020; v1 submitted 31 January, 2020; originally announced February 2020.

A lithium-ion battery based on a graphene nanoflakes ink anode and a lithium iron phosphate cathode

Authors: Jusef Hassoun, Francesco Bonaccorso, Marco Agostini, Marco Angelucci, Maria Grazia Betti, Roberto Cingolani, Mauro Gemmi, Carlo Mariani, Stefania Panero, Vittorio Pellegrini, Bruno Scrosati

Abstract: Li-ion rechargeable batteries have enabled the wireless revolution transforming global communication. Future challenges, however, demands distributed energy supply at a level that is not feasible with the current energy-storage technology. New materials, capable of providing higher energy density are needed. Here we report a new class of lithium-ion batteries based on a graphene ink anode and a li… ▽ More Li-ion rechargeable batteries have enabled the wireless revolution transforming global communication. Future challenges, however, demands distributed energy supply at a level that is not feasible with the current energy-storage technology. New materials, capable of providing higher energy density are needed. Here we report a new class of lithium-ion batteries based on a graphene ink anode and a lithium iron phosphate cathode. By carefully balancing the cell composition and suppressing the initial irreversible capacity of the anode, we demonstrate an optimal battery performance in terms of specific capacity, i.e. 165 mAhg-1, estimated energy density of about 190 Whkg-1 and life, with a stable operation for over 80 charge-discharge cycles. We link these unique properties to the graphene nanoflake anode displaying crystalline order and high uptake of lithium at the edges, as well as to its structural and morphological optimization in relation to the overall battery composition. Our approach, compatible with any printing technologies, is cheap and scalable and opens up new opportunities for the development of high-capacity Li-ion batteries. △ Less

Submitted 10 March, 2014; originally announced March 2014.

Comments: 17 pages, 10 figures