MERLOT: Architectural Support for Energy-Efficient Real-Time Processing in GPUs

Santriaji, Muhammad Husni; Hoffmann, Henry

doi:10.1109/rtas.2018.00030

Cited by 11 publications

(3 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Future work would explore additional machione learning techniques to configure the floating point usage differently for different functions in the program [19], [39], [58], [59]. Another promising line of work is using a runtime system to dynamically tune floating point usage to maintain either energy or accracy constraints in a changing workload [6], [26]- [28], [34], [38], [40], [41], [52], [57], [78], or possibly implementing this control scheme in hardware [67], [68], [83].…”

Section: Outputsmentioning

confidence: 99%

NEAT: A Framework for Automated Exploration of Floating Point Approximations

Barati,

Ehudin,

Hoffmann

2021

Preprint

View full text Add to dashboard Cite

Much recent research is devoted to exploring tradeoffs between computational accuracy and energy efficiency at different levels of the system stack. Approximation at the floating point unit (FPU) allows saving energy by simply reducing the number of computed floating point bits in return for accuracy loss. Although, finding the most energy efficient approximation for various applications with minimal effort is the main challenge. To address this issue, we propose NEAT: a pin tool that helps users automatically explore the accuracy-energy tradeoff space induced by various floating point implementations. NEAT helps programmers explore the effects of simultaneously using multiple floating point implementations to achieve the lowest energy consumption for an accuracy constraint or vice versa. NEAT accepts one or more user-defined floating point implementations and programmable placement rules for where/when to apply them. NEAT then automatically replaces floating point operations with different implementations based on the user-specified rules during the runtime and explores the resulting tradeoff space to find the best use of approximate floating point implementations for the precision tuning throughout the program. We evaluate NEAT by enforcing combinations of 24/53 different floating point implementations with three sets of placement rules on a wide range of benchmarks. We find that heuristic precision tuning at the function level provides up to 22% and 48% energy savings at 1% and 10% accuracy loss comparing to applying a single implementation for the whole application. Also, NEAT is applicable to neural networks where it finds the optimal precision level for each layer considering an accuracy target for the model.

show abstract

Section: Outputsmentioning

confidence: 99%

NEAT: A Framework for Automated Exploration of Floating Point Approximations

Barati,

Ehudin,

Hoffmann

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…In [2], we assume that the WCET of an application is obtained through empirical tests. Recently, Santriaji et al [73] have proposed a hardware-based resource manager for GPUs that enforces timing guarantees with minimal energy.…”

Section: Real-time Processing On Embedded Gpusmentioning

confidence: 99%

System-Level Design of GPU-Based Embedded Systems

Maghazeh

2018

Linköping Studies in Science and Technology. Dissertations

View full text Add to dashboard Cite

Modern embedded systems deploy several hardware accelerators, in a heterogeneous manner, to deliver high-performance computing. Among such devices, graphics processing units (GPUs) have earned a prominent position by virtue of their immense computing power. However, a system design that relies on the sheer throughput of GPUs is often incapable of satisfying the strict power-and time-related constraints faced by the embedded systems. This thesis presents several system-level software techniques to optimize the design of GPU-based embedded systems under various graphics and nongraphics applications. As compared to the conventional application-level optimizations, the system-wide view of our proposed techniques brings about several advantages: First, it allows for fully incorporating the limitations and requirements of the various system parts in the design process. Second, it can unveil optimization opportunities by exposing the information flow between the processing components. Third, the techniques are generally applicable to a wide range of applications with similar characteristics. In addition, multiple system-level techniques can be combined together or with the applicationlevel techniques to further improve performance. We begin by studying some of the unique attributes of GPU-based embedded systems and discussing several factors that distinguish the design of these systems from that of the conventional high-end GPU-based systems. We then proceed to develop two techniques that address an important challenge in the design of GPU-based embedded systems, from two different perspectives. The challenge arises from the fact that GPUs require a large amount of workload to be present at runtime in order to deliver a high throughput. However, for some embedded applications, collecting large batches of input data requires an unacceptable waiting time, prompting a trade-off between throughput and latency. We also develop an optimization technique for GPU-based applications to address the memory bottleneck issue by utilizing the GPU L2 cache to shorten data access time. Moreover, in the area of graphics applications, with a focus on mobile games, we propose a power management scheme to reduce the GPU power consumption by dynamically adjusting the display resolution while considering the user's visual perception at various resolutions. We also discuss the collective impact of the proposed techniques in tackling the design challenges of the emerging complex systems. The proposed techniques are assessed by real-life experimentations on GPU-based hardware platforms, which demonstrate the superior performance of our approaches as compared to the state-of-the-art techniques. The research presented in this thesis has been partially funded by the National Computer Science Graduate School (cugs) in Sweden. v First and foremost, I would like to thank Professor Zebo Peng, Professor Petru Eles, and Dr. Unmesh Bordoloi for their excellent support and thoughtful supervision of my Ph.D. studies and for their patience and understanding...

show abstract

“…The major practical shortcoming of our method is that: (1) It requires a large amount of manually annotated dataset to achieve satisfactory recognition results. (2) The two tasks of segmentation and recognition often use multiple models, resulting in increased data interaction between the CPU and GPU, inefficient resource allocation, and reduced recognition efficiency [2] . Since the character types are limited and their features, such as shape and color, are stable, data augmentation can be used to rapidly expand the dataset using smallsample data.…”

Section: Introductionmentioning

confidence: 99%