skip to main content

Title: BranchNet: A Convolutional Neural Network to Predict Hard-To-Predict Branches
Abstract—The state-of-the-art branch predictor, TAGE, re- mains inefficient at identifying correlated branches deep in a noisy global branch history. We argue this inefficiency is a fundamental limitation of runtime branch prediction and not a coincidental artifact due to the design of TAGE. To further improve branch prediction, we need to relax the constraint of runtime only training and adopt more sophisticated prediction mechanisms. To this end, Tarsa et al. proposed using convo- lutional neural networks (CNNs) that are trained at compile- time to accurately predict branches that TAGE cannot. Given enough profiling coverage, CNNs learn input-independent branch correlations that can accurately predict branches when running a program with unseen inputs. We build on their work and introduce BranchNet, a CNN with a practical on-chip inference engine tailored to the needs of branch prediction. At runtime, BranchNet predicts a few hard-to-predict branches, while TAGE- SC-L predicts the remaining branches. This hybrid approach reduces the MPKI of SPEC2017 Integer benchmarks by 7.6% (and up to 15.7%) when compared to a very large (impractical) MTAGE-SC baseline, demonstrating a fundamental advantage in the prediction capabilities of BranchNet compared to TAGE- like predictors. We also propose a practical resource-constrained variant of BranchNet that improves the MPKI by 9.6% (and up to 17.7%) compared to a 64KB TAGE-SC-L without increasing the prediction latency.
Authors:
; ; ;
Award ID(s):
2011145
Publication Date:
NSF-PAR ID:
10249269
Journal Name:
53rd Annual IEEE/ACM International Symposium on Microarchitecture
Page Range or eLocation-ID:
118 to 130
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract—Multi-layer neural networks show promise in im- proving branch prediction accuracy. Tarsa et al. have shown that convolutional neural networks (CNNs) can accurately predict many branches that state-of-the-art branch predictors cannot. Yet, strict latency and storage constraints make naive adoption of typical neural network architectures impractical. Thus, it is necessary to understand the unique characteristics of branch prediction to design constraint-aware neural networks. This paper studies why CNNs are so effective for two hard-to- predict branches from the SPEC benchmark suite. We identify custom prediction algorithms for these branches that are more accurate and cost-efficient than CNNs. Finally, we discuss why out-of-the-box machine learning techniques do not find optimal solutions and propose research directions aimed at solving these inefficiencies.
  2. Fariselli, Piero (Ed.)
    Predicting mutation-induced changes in protein thermodynamic stability (ΔΔG) is of great interest in protein engineering, variant interpretation, and protein biophysics. We introduce ThermoNet, a deep, 3D-convolutional neural network (3D-CNN) designed for structure-based prediction of ΔΔGs upon point mutation. To leverage the image-processing power inherent in CNNs, we treat protein structures as if they were multi-channel 3D images. In particular, the inputs to ThermoNet are uniformly constructed as multi-channel voxel grids based on biophysical properties derived from raw atom coordinates. We train and evaluate ThermoNet with a curated data set that accounts for protein homology and is balanced with direct and reverse mutations; this provides a framework for addressing biases that have likely influenced many previous ΔΔG prediction methods. ThermoNet demonstrates performance comparable to the best available methods on the widely used S sym test set. In addition, ThermoNet accurately predicts the effects of both stabilizing and destabilizing mutations, while most other methods exhibit a strong bias towards predicting destabilization. We further show that homology between S sym and widely used training sets like S2648 and VariBench has likely led to overestimated performance in previous studies. Finally, we demonstrate the practical utility of ThermoNet in predicting the ΔΔGs for two clinicallymore »relevant proteins, p53 and myoglobin, and for pathogenic and benign missense variants from ClinVar. Overall, our results suggest that 3D-CNNs can model the complex, non-linear interactions perturbed by mutations, directly from biophysical properties of atoms.« less
  3. Abstract We evaluate Beer–Lambert (BL) ray-tracing and Monte Carlo N-Particle (MCNP) photon tracking simulations for prediction and comparison of x-ray imaging system performance. These simulation tools can aid the methodical design of laboratory-scale x-ray particle image velocimetry (XPIV) experiments and tracer particles by predicting image quality. Particle image signal-to-noise ratio (SNR) is used as the metric of system performance. Simulated and experiment data of hollow, silver-coated, glass sphere tracer particles (AGSF-33) are compared. As predicted by the simulations, the AGSF-33 particles are visible with a SNR greater than unity in 100 ms exposure time images, demonstrating their potential as x-ray PIV or particle tracking velocimetry (XPTV) tracers. The BL approach predicts the image contrast, is computationally inexpensive, and enables the exploration of a vast parameter space for system design. MCNP simulations, on the other hand, predict experiment images slightly more accurately, but are more than an order of magnitude more computationally expensive than BL simulations. For most practical XPIV system design applications, the higher computational expense of MCNP is likely not justified by the modest accuracy improvement compared to BL.
  4. Structures with adaptive capabilities offer many potentials to achieve future needs in efficiency, reliability, and intelligence. To this end, bistable CFRP (Carbon Fibre Reinforced Polymers) composites with asymmetric fiber layout are a promising concept that has shown shape morphing capabilities that adapt to the changes in the environment such as external forces and moments. This adaptability opens them to endless application potentials, ranging from small micro-switches to large airfoil sections in airplane wings or wind turbine blades. To harness this potential, it is essential to predict these composites’ physical shapes and behavior accurately. To this end, Hyer and Dano devised the first analytical model based on the concepts of Classical Lamination Theory, and this model has become the cornerstone of almost all subsequent studies. However, this theory uses Kirchoff’s theory of thin plates that are limited by several assumptions. As a result, Hyer’s theory can predict the overall shape of these laminates but lacks accuracy. A reason for this model’s underperformance is that it ignores the inter-laminar stresses and strains, but such stresses/strains play a vital role in the balance of the overall stress field and are found significantly higher near the free edges. To overcome these fundamental limitations, wemore »propose a new analytical approach by combining the Reissner-Mindlin theory with concepts from the Classical Lamination Theory. This new model introduces in-plane rotations as two additional degrees of freedom. Thus, it has five independent variables compared to only three in Hyer and Dano’s model and its derivatives. Hence, we have a more complex but more accurate model. This paper outlines our new analytical approach by 1) introducing these two additional degrees of freedom; 2) selecting appropriate polynomial approximations; 3) formulating inter-laminar stresses that are functions of these added rotations; and 4) incorporating these inter-laminar stresses in the potential energy equation. By comparing this model’s prediction with the finite element simulation results, we found the new model slightly under predicts the laminate deformation, but the overall accuracy is promising, as evidenced by high R-squared correlation.« less
  5. Abstract

    Access to accurate, generalizable and scalable solar irradiance prediction is critical for smooth solar-grid integration, especially in the light of the accelerated global adoption of solar energy production. Both physical and statistical prediction models of solar irradiance have been proposed in the literature. Physical models require meteorological forecasts—generated by computationally expensive models—to predict solar irradiance, with limited accuracy in sub-daily predictions. Statistical models leveragein-situmeasurements which require expensive equipment and do not account for meso-scale atmospheric dynamics. We address these fundamental gaps by developing a convolutional global horizontal irradiance prediction model, using convolutional neural networks and publicly accessible satellite cloud images. Our proposed model predicts solar irradiance in 12 different locations in the US for various prediction time horizons. Our model yields up to 24% improvement in an hour-ahead predictions and 26% in a day-ahead predictions compared to a persistence forecast. Moreover, using saliency maps and target-location-focused cropping, we demonstrate the benefits of incorporating meso-scale atmospheric dynamics for prediction performance. Our results are critical for energy systems planners, utility managers and electricity market participants to ensure efficient harvesting of the solar energy and reliable operation of the grid.