-
Guac: Energy-Aware and SSA-Based Generation of Coarse-Grained Merged Accelerators from LLVM-IR
Authors:
Iulian Brumar,
Rodrigo Rocha,
Alex Bernat,
Devashree Tripathy,
David Brooks,
Gu-Yeon Wei
Abstract:
Designing accelerators for resource- and power-constrained applications is a daunting task. High-level Synthesis (HLS) addresses these constraints through resource sharing, an optimization at the HLS binding stage that maps multiple operations to the same functional unit.
However, resource sharing is often limited to reusing instructions within a basic block. Instead of searching globally for th…
▽ More
Designing accelerators for resource- and power-constrained applications is a daunting task. High-level Synthesis (HLS) addresses these constraints through resource sharing, an optimization at the HLS binding stage that maps multiple operations to the same functional unit.
However, resource sharing is often limited to reusing instructions within a basic block. Instead of searching globally for the best control and dataflow graphs (CDFGs) to combine, it is constrained by existing instruction map**s and schedules.
Coarse-grained function merging (CGFM) at the intermediate representation (IR) level can reuse control and dataflow patterns without dealing with the post-scheduling complexity of map** operations onto functional units, wires, and registers. The merged functions produced by CGFM can be translated to RTL by HLS, yielding Coarse Grained Merged Accelerators (CGMAs). CGMAs are especially profitable across applications with similar data- and control-flow patterns. Prior work has used CGFM to generate CGMAs without regard for which CGFM algorithms best optimize area, power, and energy costs.
We propose Guac, an energy-aware and SSA-based (static single assignment) CGMA generation methodology. Guac implements a novel ensemble of cost models for efficient CGMA generation. We also show that CGFM algorithms using SSA form to merge control- and dataflow graphs outperform prior non-SSA CGFM designs. We demonstrate significant area, power, and energy savings with respect to the state of the art. In particular, Guac more than doubles energy savings with respect to the closest related work while using a strong resource-sharing baseline.
△ Less
Submitted 20 February, 2024;
originally announced February 2024.
-
Quantum information scrambling in two-dimensional Bose-Hubbard lattices
Authors:
Devjyoti Tripathy,
Akram Touil,
Bartłomiej Gardas,
Sebastian Deffner
Abstract:
It is a well-understood fact that the transport of excitations throughout a lattice is intimately governed by the underlying structures. Hence, it is only natural to recognize that also the dispersion of information has to depend on the lattice geometry. In the present work, we demonstrate that two-dimensional lattices described by the Bose-Hubbard model exhibit information scrambling for systems…
▽ More
It is a well-understood fact that the transport of excitations throughout a lattice is intimately governed by the underlying structures. Hence, it is only natural to recognize that also the dispersion of information has to depend on the lattice geometry. In the present work, we demonstrate that two-dimensional lattices described by the Bose-Hubbard model exhibit information scrambling for systems as little as two hexagons. However, we also find that the OTOC shows the exponential decay characteristic for quantum chaos only for a judicious choice of local observables. More generally, the OTOC is better described by Gaussian-exponential convolutions, which alludes to the close similarity of information scrambling and decoherence theory.
△ Less
Submitted 16 January, 2024;
originally announced January 2024.
-
ArchGym: An Open-Source Gymnasium for Machine Learning Assisted Architecture Design
Authors:
Srivatsan Krishnan,
Amir Yazdanbaksh,
Shvetank Prakash,
Jason Jabbour,
Ikechukwu Uchendu,
Susobhan Ghosh,
Behzad Boroujerdian,
Daniel Richins,
Devashree Tripathy,
Aleksandra Faust,
Vijay Janapa Reddi
Abstract:
Machine learning is a prevalent approach to tame the complexity of design space exploration for domain-specific architectures. Using ML for design space exploration poses challenges. First, it's not straightforward to identify the suitable algorithm from an increasing pool of ML methods. Second, assessing the trade-offs between performance and sample efficiency across these methods is inconclusive…
▽ More
Machine learning is a prevalent approach to tame the complexity of design space exploration for domain-specific architectures. Using ML for design space exploration poses challenges. First, it's not straightforward to identify the suitable algorithm from an increasing pool of ML methods. Second, assessing the trade-offs between performance and sample efficiency across these methods is inconclusive. Finally, lack of a holistic framework for fair, reproducible, and objective comparison across these methods hinders progress of adopting ML-aided architecture design space exploration and impedes creating repeatable artifacts. To mitigate these challenges, we introduce ArchGym, an open-source gym and easy-to-extend framework that connects diverse search algorithms to architecture simulators. To demonstrate utility, we evaluate ArchGym across multiple vanilla and domain-specific search algorithms in designing custom memory controller, deep neural network accelerators, and custom SoC for AR/VR workloads, encompassing over 21K experiments. Results suggest that with unlimited samples, ML algorithms are equally favorable to meet user-defined target specification if hyperparameters are tuned; no solution is necessarily better than another (e.g., reinforcement learning vs. Bayesian methods). We coin the term hyperparameter lottery to describe the chance for a search algorithm to find an optimal design provided meticulously selected hyperparameters. The ease of data collection and aggregation in ArchGym facilitates research in ML-aided architecture design space exploration. As a case study, we show this advantage by develo** a proxy cost model with an RMSE of 0.61% that offers a 2,000-fold reduction in simulation time. Code and data for ArchGym is available at https://bit.ly/ArchGym.
△ Less
Submitted 15 June, 2023;
originally announced June 2023.
-
PerfSAGE: Generalized Inference Performance Predictor for Arbitrary Deep Learning Models on Edge Devices
Authors:
Yuji Chai,
Devashree Tripathy,
Chuteng Zhou,
Dibakar Gope,
Igor Fedorov,
Ramon Matas,
David Brooks,
Gu-Yeon Wei,
Paul Whatmough
Abstract:
The ability to accurately predict deep neural network (DNN) inference performance metrics, such as latency, power, and memory footprint, for an arbitrary DNN on a target hardware platform is essential to the design of DNN based models. This ability is critical for the (manual or automatic) design, optimization, and deployment of practical DNNs for a specific hardware deployment platform. Unfortuna…
▽ More
The ability to accurately predict deep neural network (DNN) inference performance metrics, such as latency, power, and memory footprint, for an arbitrary DNN on a target hardware platform is essential to the design of DNN based models. This ability is critical for the (manual or automatic) design, optimization, and deployment of practical DNNs for a specific hardware deployment platform. Unfortunately, these metrics are slow to evaluate using simulators (where available) and typically require measurement on the target hardware. This work describes PerfSAGE, a novel graph neural network (GNN) that predicts inference latency, energy, and memory footprint on an arbitrary DNN TFlite graph (TFL, 2017). In contrast, previously published performance predictors can only predict latency and are restricted to pre-defined construction rules or search spaces. This paper also describes the EdgeDLPerf dataset of 134,912 DNNs randomly sampled from four task search spaces and annotated with inference performance metrics from three edge hardware platforms. Using this dataset, we train PerfSAGE and provide experimental results that demonstrate state-of-the-art prediction accuracy with a Mean Absolute Percentage Error of <5% across all targets and model search spaces. These results: (1) Outperform previous state-of-art GNN-based predictors (Dudziak et al., 2020), (2) Accurately predict performance on accelerators (a shortfall of non-GNN-based predictors (Zhang et al., 2021)), and (3) Demonstrate predictions on arbitrary input graphs without modifications to the feature extractor.
△ Less
Submitted 26 January, 2023;
originally announced January 2023.
-
Bootstrap** PT symmetric Hamiltonians
Authors:
Sakil Khan,
Yuv Agarwal,
Devjyoti Tripathy,
Sachin Jain
Abstract:
Bootstrap** in Quantum Mechanics uses positivity condition to derive the Eigenspectum. For non-hermitian systems usual positivity condition does not work. In this paper we define positivity condition for special class of non-hermitian hamiltonian, the PT symmetric Hamiltonian. We illustrate this modified positivity condition with several examples and obtain eigenspectrum.
Bootstrap** in Quantum Mechanics uses positivity condition to derive the Eigenspectum. For non-hermitian systems usual positivity condition does not work. In this paper we define positivity condition for special class of non-hermitian hamiltonian, the PT symmetric Hamiltonian. We illustrate this modified positivity condition with several examples and obtain eigenspectrum.
△ Less
Submitted 10 February, 2022;
originally announced February 2022.
-
Raman and first-principles study of the pressure induced Mott-insulator to metal transition in bulk FePS$_3$
Authors:
Subhadip Das,
Shashank Chaturvedi,
Debashis Tripathy,
Shivani Grover,
Rajendra Singh,
D. V. S. Muthu,
S. Sampath,
U. V. Waghmare,
A. K. Sood
Abstract:
Recently discovered class of 2D materials based on transition metal phosphorous trichalcogenides exhibit antiferromagnetic ground state, with potential applications in spintronics. Amongst them, FePS$ _{3} $ is a Mott insulator with a band gap of $\sim$ 1.5 eV. This study using Raman spectroscopy along with first-principles density functional theoretical analysis examines the stability of its stru…
▽ More
Recently discovered class of 2D materials based on transition metal phosphorous trichalcogenides exhibit antiferromagnetic ground state, with potential applications in spintronics. Amongst them, FePS$ _{3} $ is a Mott insulator with a band gap of $\sim$ 1.5 eV. This study using Raman spectroscopy along with first-principles density functional theoretical analysis examines the stability of its structure and electronic properties under pressure. Raman spectroscopy reveals two phase transitions at 4.6 GPa and 12 GPa marked by the changes in pressure coefficients of the mode frequencies and the number of symmetry allowed modes. FePS$_3$ transforms from the ambient monoclinic C2/m phase with a band gap of 1.54 eV to another monoclinic C2/m (band gap of 0.1 eV) phase at 4.6 GPa, followed by another transition at 12 GPa to the metallic trigonal P-31m phase. Our work complements recently reported high pressure X-ray diffraction studies.
△ Less
Submitted 10 September, 2021;
originally announced September 2021.
-
Interaction of (-)-epigallocatechin gallate with silver nanoparticles
Authors:
Goutam Kumar Chandra,
Debi Ranjan Tripathy,
Swagata Dasgupta,
Anushree Roy
Abstract:
Interactions between silver nanoparticles and (-)-epigallocatechin gallate (EGCG) have been investigated. Prior to the addition of EGCG molecules the silver particles are stabilized by borate ions. Studies on the surface plasmon resonance band of silver particles suggest that the EGCG molecules remove the borate ions from the surface of the metal particles due to the chelating property of the io…
▽ More
Interactions between silver nanoparticles and (-)-epigallocatechin gallate (EGCG) have been investigated. Prior to the addition of EGCG molecules the silver particles are stabilized by borate ions. Studies on the surface plasmon resonance band of silver particles suggest that the EGCG molecules remove the borate ions from the surface of the metal particles due to the chelating property of the ions. The complex formation by EGCG and borate ions has been confirmed by NMR studies and pH titration. A possible scheme of interaction between the two has been proposed.
△ Less
Submitted 2 January, 2010;
originally announced January 2010.
-
A quantum mechanical derivation of Gamow's relation for the time and temperature of the expanding Universe
Authors:
S. Mishra,
D. N. Tripathy
Abstract:
The quantum mechanical approach developed by us recently for the evolution of the universe is used to derive an alternative derivation connecting the temperature of the cosmic background radiation and the age of the universe which is found to be similar to the one obtained by Gamow long back. By assuming the age of the universe to be $\approx$ 20 billion years, we reproduce a value of $\approx$…
▽ More
The quantum mechanical approach developed by us recently for the evolution of the universe is used to derive an alternative derivation connecting the temperature of the cosmic background radiation and the age of the universe which is found to be similar to the one obtained by Gamow long back. By assuming the age of the universe to be $\approx$ 20 billion years, we reproduce a value of $\approx$ 2.91 K for the cosmic back-ground radiation, agreeing well with the recently measured experimental value of 2.728 K. Besides, this theory enables us to calculate the photon density and entropy associated with the background radiation and the ratio of the number of photons to the number of nucleons, which quantitatively agree with the results obtained by others.
△ Less
Submitted 10 December, 1999;
originally announced December 1999.
-
An Understanding of The Dark Matter in The Universe And The Variation of The Universal Gravitational Constant G With Time
Authors:
D. N. Tripathy,
Subodha Mishra
Abstract:
Considering the fact that the present universe might have been formed out of a system of ficticious self-gravitating particles, fermionic in nature, each of mass $m$, we are able to obtain a compact expression for the radius $R_0$ of the universe by using a model density distribution $ρ(r)$ for the particles which is singular at the origin. This singularity in $ρ(r)$ can be considered to be cons…
▽ More
Considering the fact that the present universe might have been formed out of a system of ficticious self-gravitating particles, fermionic in nature, each of mass $m$, we are able to obtain a compact expression for the radius $R_0$ of the universe by using a model density distribution $ρ(r)$ for the particles which is singular at the origin. This singularity in $ρ(r)$ can be considered to be consistent with the socalled Big Bang theory of the universe. By assuming that Mach's principle holds good in the evolution of the universe, we determine the number of particles, $N$, of the universe and its $R_0$, which are obtained in terms of the mass $m$ of the constituent particles and the Universal Gravitational constant $G$ only. It is seen that for a mass of the constituent particles $m\simeq 1.07\times 10^{-35} g$ the age of the present universe,$τ_0$, becomes $τ_0 \simeq 20\times 10^9 yr$, or equivalently $R_0 \simeq 1.9\times 10^{28} cm $. For this $m$, the total number of particles costituting the present universe is found to be $N \simeq 2.4 \times 10^{91}$ and its total mass $(M \simeq 1.27916 \times 10^{23} M_{\odot})$, $M_{\odot}$ being the solar mass. All these numbers seem to be quantitatively agreeing with those evaluated from other theories. Using the present theory, we have also made an estimation of the variation of the universal gravitational constant $G$ with time which gives $({\dot G \over G}) =-9.6\times 10^{-11} yr^{-1}$. This is again in extremely good agreement with the results of some of the most recent calculations. Lastly, a plausible explanation for the Dark Matter present in today's universe is given.
△ Less
Submitted 12 May, 1997; v1 submitted 11 May, 1997;
originally announced May 1997.
-
A Quantum Mechanical Approach To A System of Self-Gravitating Particles And The Problem of Gravitational Collapse
Authors:
D. N. Tripathy,
Subodha Mishra
Abstract:
By making an intuitive choice for the single-particle density of a system of N self-gravitating particles, without any source for the radiation of energy, we have been able to calculate the binding energy of the system by treating these particles as fermions. Our expression for the ground state energy of the system shows a dependence of $N^{7/3}$ on the particle number, which is in agreement wit…
▽ More
By making an intuitive choice for the single-particle density of a system of N self-gravitating particles, without any source for the radiation of energy, we have been able to calculate the binding energy of the system by treating these particles as fermions. Our expression for the ground state energy of the system shows a dependence of $N^{7/3}$ on the particle number, which is in agreement with the results obtained by other workers. We also arrive at a compact expression for the radius of a star following which we correctly reproduce the nucleon number to be found in a typical star. Using this value, we obtain the well-known result for the limiting value of the mass, M, of a neutron star $(M \simeq 3.12 M_{\odot}, M_{\odot}$ being the solar mass) beyond which the black hole formation should take place. Generalizing the present calculation to the case of white dwarfs,we have been able to obtain the so called Chandrasekhar limit for the mass, $M_{Ch}$, $(M_{Ch}\simeq 1.44 M_{\odot})$ below which the stars are expected to go over to the white dwarf state. We reproduce this by introducing a radius, equivalent to Schwarzschild radius, at the interface of the neutron stars and white dwarfs. This is justified by considering the fact that it gives rise to the correct value for the degree of ionization $μ_e (μ_e\approx 2)$ for heavy nuclei.
△ Less
Submitted 10 December, 1996;
originally announced December 1996.
