Showing 1–7 of 7 results for author: Wilkening, M

2D Li$^{\bf +}$ ionic hop** in Li$_{\bf 3}$InCl$_{\bf 6}$ as revealed by diffusion-induced nuclear spin relaxation

Authors: Florian Stainer, H. Martin R. Wilkening

Abstract: Ternary Li halides, such as Li$_3$MeX$_6$ with, e.g., Me = In, Sc, Y and X = Cl, Br, are in the center of attention for battery applications as these materials might serve as ionic electrolytes. To fulfill their function, such electrolytes must have an extraordinarily high ionic Li$^+$ conductivity. Layer-structured Li$_3$InCl$_6$ represents such a candidate; understanding the origin of the rapid… ▽ More Ternary Li halides, such as Li$_3$MeX$_6$ with, e.g., Me = In, Sc, Y and X = Cl, Br, are in the center of attention for battery applications as these materials might serve as ionic electrolytes. To fulfill their function, such electrolytes must have an extraordinarily high ionic Li$^+$ conductivity. Layer-structured Li$_3$InCl$_6$ represents such a candidate; understanding the origin of the rapid Li$^+$ exchange processes needs, however, further investigation. Spatially restricted, that is, low-dimensional particle diffusion might offer an explanation for fast ion dynamics. It is, however, challenging to provide evidence for 2D diffusion at the atomic scale when dealing with polycrystalline powder samples. Here, we used purely diffusion-induced $^7$Li nuclear magnetic spin relaxation to detect anomalies that unambiguously show that 2D Li diffusion is chiefly responsible for the dynamic processes in a Li$_3$InCl$_6$ powder sample. The change of the spin-lattice relaxation rate $1/T_1$ as a function of inverse temperature $1/T$ passes through a rate peak that is strictly following asymmetric behavior. This feature is in excellent agreement with the model of P. M. Richards suggesting a logarithmic spectral density function $J$ to fully describe 2D diffusion. Hence, Li$_3$InCl$_6$ belongs to the very rare examples for which 2D Li$^+$ diffusion has been immaculately verified. We believe that such information help understand the dynamic features of ternary Li halides. △ Less

Submitted 11 January, 2024; originally announced January 2024.

RecPipe: Co-designing Models and Hardware to Jointly Optimize Recommendation Quality and Performance

Authors: Udit Gupta, Samuel Hsia, Jeff Zhang, Mark Wilkening, Javin Pombra, Hsien-Hsin S. Lee, Gu-Yeon Wei, Carole-Jean Wu, David Brooks

Abstract: Deep learning recommendation systems must provide high quality, personalized content under strict tail-latency targets and high system loads. This paper presents RecPipe, a system to jointly optimize recommendation quality and inference performance. Central to RecPipe is decomposing recommendation models into multi-stage pipelines to maintain quality while reducing compute complexity and exposing… ▽ More Deep learning recommendation systems must provide high quality, personalized content under strict tail-latency targets and high system loads. This paper presents RecPipe, a system to jointly optimize recommendation quality and inference performance. Central to RecPipe is decomposing recommendation models into multi-stage pipelines to maintain quality while reducing compute complexity and exposing distinct parallelism opportunities. RecPipe implements an inference scheduler to map multi-stage recommendation engines onto commodity, heterogeneous platforms (e.g., CPUs, GPUs).While the hardware-aware scheduling improves ranking efficiency, the commodity platforms suffer from many limitations requiring specialized hardware. Thus, we design RecPipeAccel (RPAccel), a custom accelerator that jointly optimizes quality, tail-latency, and system throughput. RPAc-cel is designed specifically to exploit the distinct design space opened via RecPipe. In particular, RPAccel processes queries in sub-batches to pipeline recommendation stages, implements dual static and dynamic embedding caches, a set of top-k filtering units, and a reconfigurable systolic array. Com-pared to prior-art and at iso-quality, we demonstrate that RPAccel improves latency and throughput by 3x and 6x. △ Less

Submitted 22 May, 2021; v1 submitted 18 May, 2021; originally announced May 2021.

RecSSD: Near Data Processing for Solid State Drive Based Recommendation Inference

Authors: Mark Wilkening, Udit Gupta, Samuel Hsia, Caroline Trippel, Carole-Jean Wu, David Brooks, Gu-Yeon Wei

Abstract: Neural personalized recommendation models are used across a wide variety of datacenter applications including search, social media, and entertainment. State-of-the-art models comprise large embedding tables that have billions of parameters requiring large memory capacities. Unfortunately, large and fast DRAM-based memories levy high infrastructure costs. Conventional SSD-based storage solutions of… ▽ More Neural personalized recommendation models are used across a wide variety of datacenter applications including search, social media, and entertainment. State-of-the-art models comprise large embedding tables that have billions of parameters requiring large memory capacities. Unfortunately, large and fast DRAM-based memories levy high infrastructure costs. Conventional SSD-based storage solutions offer an order of magnitude larger capacity, but have worse read latency and bandwidth, degrading inference performance. RecSSD is a near data processing based SSD memory system customized for neural recommendation inference that reduces end-to-end model inference latency by 2X compared to using COTS SSDs across eight industry-representative models. △ Less

Submitted 29 January, 2021; originally announced February 2021.

Cross-Stack Workload Characterization of Deep Recommendation Systems

Authors: Samuel Hsia, Udit Gupta, Mark Wilkening, Carole-Jean Wu, Gu-Yeon Wei, David Brooks

Abstract: Deep learning based recommendation systems form the backbone of most personalized cloud services. Though the computer architecture community has recently started to take notice of deep recommendation inference, the resulting solutions have taken wildly different approaches - ranging from near memory processing to at-scale optimizations. To better design future hardware systems for deep recommendat… ▽ More Deep learning based recommendation systems form the backbone of most personalized cloud services. Though the computer architecture community has recently started to take notice of deep recommendation inference, the resulting solutions have taken wildly different approaches - ranging from near memory processing to at-scale optimizations. To better design future hardware systems for deep recommendation inference, we must first systematically examine and characterize the underlying systems-level impact of design decisions across the different levels of the execution stack. In this paper, we characterize eight industry-representative deep recommendation models at three different levels of the execution stack: algorithms and software, systems platforms, and hardware microarchitectures. Through this cross-stack characterization, we first show that system deployment choices (i.e., CPUs or GPUs, batch size granularity) can give us up to 15x speedup. To better understand the bottlenecks for further optimization, we look at both software operator usage breakdown and CPU frontend and backend microarchitectural inefficiencies. Finally, we model the correlation between key algorithmic model architecture features and hardware bottlenecks, revealing the absence of a single dominant algorithmic component behind each hardware bottleneck. △ Less

Submitted 10 October, 2020; originally announced October 2020.

Comments: Published in 2020 IEEE International Symposium on Workload Characterization (IISWC)

Singlet Oxygen Generation as a Major Cause for Parasitic Reactions during Cycling of Aprotic Lithium-Oxygen Batteries

Authors: Nika Mahne, Bettina Schafzahl, Christian Leypold, Mario Leypold, Sandra Grumm, Anita Leitgeb, Gernot A. Strohmeier, Martin Wilkening, Olivier Fontaine, Denis Kramer, Christian Slugovc, Sergey M. Borisov, Stefan A. Freunberger

Abstract: Non-aqueous metal-oxygen batteries depend critically on the reversible formation/decomposition of metal oxides on cycling. Irreversible parasitic reactions cause poor rechargeability, efficiency, and cycle life and have predominantly been ascribed to the reactivity of reduced oxygen species with cell components. These species, however, cannot fully explain the side reactions. Here we show that sin… ▽ More Non-aqueous metal-oxygen batteries depend critically on the reversible formation/decomposition of metal oxides on cycling. Irreversible parasitic reactions cause poor rechargeability, efficiency, and cycle life and have predominantly been ascribed to the reactivity of reduced oxygen species with cell components. These species, however, cannot fully explain the side reactions. Here we show that singlet oxygen forms at the cathode of a lithium-oxygen cell during discharge and from the onset of charge, and accounts for the majority of parasitic reaction products. The amount increases during discharge, early stages of charge, and charging at higher voltages, and is enhanced by the presence of trace water. Superoxide and peroxide appear to be involved in singlet oxygen generation. Singlet oxygen traps and quenchers can reduce parasitic reactions effectively. Awareness of the highly reactive singlet oxygen in non-aqueous metal-oxygen batteries gives a rationale for future research towards achieving highly reversible cell operation. △ Less

Submitted 28 November, 2017; originally announced November 2017.

Journal ref: Nature Energy 2, 17036 (2017)

$\rm\bf {^{23}Na}$ NMR spin-lattice relaxation reveals ultrafast $\rm\bf Na{^{+}}$ ion dynamics in the solid electrolyte $\rm\bf Na{_{3.4}}Sc{_{0.4}}Zr{_{1.6}}(SiO{_{4}}){_{2}}PO{_{4}}$

Authors: S. Lunghammer, D. Prutsch, S. Breuer, D. Rettenwander, I. Hanzu, Q. Ma, F. Tietz, M. Wilkening

Abstract: The realization of green and economically friendly energy storage systems needs materials with outstanding properties. Future batteries based on Na as an abundant element take advantage of non-flammable ceramic electrolytes with very high conductivities. $\rm Na{_{3}}Zr{_{2}}(SiO{_{4}}){_{2}}PO{_{4}}$-type superionic conductors are expected to pave the way for inherently safe and sustainable all-s… ▽ More The realization of green and economically friendly energy storage systems needs materials with outstanding properties. Future batteries based on Na as an abundant element take advantage of non-flammable ceramic electrolytes with very high conductivities. $\rm Na{_{3}}Zr{_{2}}(SiO{_{4}}){_{2}}PO{_{4}}$-type superionic conductors are expected to pave the way for inherently safe and sustainable all-solid-state batteries. So far, only little information has been extracted from spectroscopic measurements to clarify the origins of fast ionic hop** on the atomic length scale. Here we combined broad-band conductivity spectroscopy and nuclear magnetic resonance (NMR) relaxation to study Na ion dynamics from the μm to the angstrom length scale. Spin-lattice relaxation NMR revealed a very fast Na ion exchange process in $\rm Na{_{3.4}}Sc{_{0.4}}Zr{_{1.6}}(SiO{_{4}}){_{2}}PO{_{4}}$ that is characterized by an unprecedentedly high self-diffusion coefficient of $\rm 9 \times 10{^{-12}} m{^{2}}s{^{-1}}$ at $\rm -10°C$. Thus, well below ambient temperature the Na ions have access to elementary diffusion processes with a mean residence time $\rm τ_{NMR}$ of only $\rm 2\; ns$. The underlying asymmetric diffusion-induced NMR rate peak and the corresponding conductivity isotherms measured in the MHz range reveal correlated ionic motion. Obviously, local but extremely rapid $\rm Na{^{+}}$ jumps, involving especially the transition sites in Sc-NZSP, trigger long-range ion transport and push ionic conductivity up to $\rm 2\; mS\; cm{^{-1}}$ at room temperature. △ Less

Submitted 11 September, 2017; originally announced September 2017.

Superionic diffusion through frustrated energy landscape

Authors: D. Di Stefano, A. Miglio, K. Robeyns, Y. Filinchuk, M. Lechartier, A. Senyshyn, H. Ishida, S. Spannenberger, D. Prutsch, S. Lunghammer, D. Rettenwander, M. Wilkening, B. Roling, Y. Kato, G. Hautier

Abstract: Solid-state materials with high ionic conduction are necessary to many technologies including all-solid-state Li-ion batteries. Understanding how crystal structure dictates ionic diffusion is at the root of the development of fast ionic conductors. Here, we show that LiTi2(PS4)3 exhibits a Li-ion diffusion coefficient about an order of magnitude higher than current state-of-the-art lithium superio… ▽ More Solid-state materials with high ionic conduction are necessary to many technologies including all-solid-state Li-ion batteries. Understanding how crystal structure dictates ionic diffusion is at the root of the development of fast ionic conductors. Here, we show that LiTi2(PS4)3 exhibits a Li-ion diffusion coefficient about an order of magnitude higher than current state-of-the-art lithium superionic conductors. We rationalize this observation by the unusual crystal structure of LiTi2(PS4)3 which offers no regular tetrahedral or octahedral sites for lithium to favorably occupy. This creates a smooth, frustrated energy landscape resembling more the energy landscapes present in liquids than in typical solids. This frustrated energy landscape leads to a high diffusion coefficient combining low activation energy with a high pre-factor. △ Less

Submitted 24 July, 2019; v1 submitted 8 August, 2017; originally announced August 2017.