02.06.2023
| Research |
A New Way Towards Performance Prediction of ML Accelerators
Machine learning algorithms like CNNs are crucial for emerging technologies such as autonomous driving, the Internet of Things (IoT) or Edge Computing. Finding the most appropriate deep learning (DL) accelerator for CNNs is essential and comes with different challenges. A newly developed method enables a fast and accurate performance prediction of CNNs for GPGPUs during the early stages of the design process.
Convolutional Neural Networks (CNNs) can handle huge amounts of unstructured data and are widely used to perform various tasks such as image and video recognition, classification, and natural language processing. Moreover, CNNs are one of the most popular deep learning algorithms used for emerging technologies such as autonomous driving. They are highly complex algorithms that consist of multiple types of layers and thus require high computational resources. Consequently, designers often use Machine learning (ML) accelerators like
General Purpose Computation on Graphics Processing Units (GPGPUs) to improve and speed-up their performance.
Hence, selecting the right ML accelerator is of great importance. This process, however, comes with challenges and can be very time-consuming as well as cost-intensive. To overcome this issue, a fast and automated approach is needed which allows finding the right accelerator without the need to build several prototypes and test different hardware platforms.
Together with Mehran Goli (DFKI) and under the direction of Rolf Drechsler, DSC research associate Christopher Metz has a developed a novel approach which enables fast and accurate performance estimation of CNNs for GPGPUs during the early stages of the design process. This new method is discussed in the paper “Fast and Accurate: Machine Learning Techniques for Performance Estimation of CNNs for GPGPUs” and is linked to previous studies of the authors which focused on the determination of GPGPUs’ power consumption (we reported
»here).
The study compared five different ML algorithms to obtain the most accurate predictive performance model. In addition, the developed approach was evaluated by estimating the performance of several CNNs for different GPGPUs (NVIDIA V100S and NVIDIA GTX 1080iT). The results show that the proposed prediction model can determine a CNNs performance with an absolute percentage error of 5,73% compared to real GPGPUs. The new method also is significantly faster than the execution time of previous approaches which enables the performance estimation of CNNs at the early stages of the design process. In contrast to many comparable approaches, this method does not require a prior execution of the neural networks. Furthermore, the model is not limited to a single GPGPU because it considers hardware features and thereby supports cross-platform prediction.
The work was presented on May 19th (4:10 pm (MEZ)) at the PAISE 2023 workshop on “Parallel AI and Systems for the Edge” in St. Petersburg, Florida. The interaction-focused workshop took place for the fifth time and was co-conducted with IPDPS 2023.
Updated by: Svenja Goers
« back
02.06.2023 | Research
A New Way Towards Performance Prediction of ML Accelerators
Machine learning algorithms like CNNs are crucial for emerging technologies such as autonomous driving, the Internet of Things (IoT) or Edge Computing. Finding the most appropriate deep learning (DL) accelerator for CNNs is essential and comes with different challenges. A newly developed method enables a fast and accurate performance prediction of CNNs for GPGPUs during the early stages of the design process.
Convolutional Neural Networks (CNNs) can handle huge amounts of unstructured data and are widely used to perform various tasks such as image and video recognition, classification, and natural language processing. Moreover, CNNs are one of the most popular deep learning algorithms used for emerging technologies such as autonomous driving. They are highly complex algorithms that consist of multiple types of layers and thus require high computational resources. Consequently, designers often use Machine learning (ML) accelerators like
General Purpose Computation on Graphics Processing Units (GPGPUs) to improve and speed-up their performance.
Hence, selecting the right ML accelerator is of great importance. This process, however, comes with challenges and can be very time-consuming as well as cost-intensive. To overcome this issue, a fast and automated approach is needed which allows finding the right accelerator without the need to build several prototypes and test different hardware platforms.
Together with Mehran Goli (DFKI) and under the direction of Rolf Drechsler, DSC research associate Christopher Metz has a developed a novel approach which enables fast and accurate performance estimation of CNNs for GPGPUs during the early stages of the design process. This new method is discussed in the paper “Fast and Accurate: Machine Learning Techniques for Performance Estimation of CNNs for GPGPUs” and is linked to previous studies of the authors which focused on the determination of GPGPUs’ power consumption (we reported
»here).
The study compared five different ML algorithms to obtain the most accurate predictive performance model. In addition, the developed approach was evaluated by estimating the performance of several CNNs for different GPGPUs (NVIDIA V100S and NVIDIA GTX 1080iT). The results show that the proposed prediction model can determine a CNNs performance with an absolute percentage error of 5,73% compared to real GPGPUs. The new method also is significantly faster than the execution time of previous approaches which enables the performance estimation of CNNs at the early stages of the design process. In contrast to many comparable approaches, this method does not require a prior execution of the neural networks. Furthermore, the model is not limited to a single GPGPU because it considers hardware features and thereby supports cross-platform prediction.
The work was presented on May 19th (4:10 pm (MEZ)) at the PAISE 2023 workshop on “Parallel AI and Systems for the Edge” in St. Petersburg, Florida. The interaction-focused workshop took place for the fifth time and was co-conducted with IPDPS 2023.
Author: Svenja Goers
Please contact us if you have any questions:
Christopher Metz
Research Associate
+49 (421) 218 - 63942
cmetz@uni-bremen.de
« back