A Survey of Three Types of Processing Units: CPU, GPU and TPU

Nikolić, Goran; Dimitrijević, Bojan; Nikolic, Tatjana R.; Stojcev, Mile K.

doi:10.1109/icest55168.2022.9828625

Cited by 10 publications

(5 citation statements)

References 3 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Python, on the other hand, is a high-level performance programming language that is widely used in machine learning (Raschka et al, 2020). TPU is a type of circuit known as ‘application-specific integrated circuits’ (ASICs) developed by Google for accelerated machine learning, and it is faster in computing than GPU (Nikolić et al, 2022). Keras is a neural network application programming interface (API) for Python that runs on top of the machine-learning platform TensorFlow (Parvat et al, 2017).…”

Section: Methodsmentioning

confidence: 99%

An artificial neural network approach for rational decision-making in borderline orthodontic cases: A preliminary analytical observational in silico study

et al. 2023

View full text Add to dashboard Cite

Introduction: Artificial intelligence (AI) technology has transformed the way healthcare functions in the present scenario. In orthodontics, expert systems and machine learning have aided clinicians in making complex, multifactorial decisions. One such scenario is an extraction decision in a borderline case. Objective: The present in silico study was planned with the intention of building an AI model for extraction decisions in borderline orthodontic cases. Design: An observational analytical study. Setting: Department of Orthodontics, Hitkarini Dental College and Hospital, Madhya Pradesh Medical University, Jabalpur, India. Methods: An artificial neural network (ANN) model for extraction or non-extraction decisions in borderline orthodontic cases was constructed based on a supervised learning algorithm using the Python (version 3.9) Sci-Kit Learn library and feed-forward backpropagation method. Based on 40 borderline orthodontic cases, 20 experienced clinicians were asked to recommend extraction or non-extraction treatment. The decision of the orthodontist and the diagnostic records, including the selected extraoral and intra-oral features, model analysis and cephalometric analysis parameters, constituted the training dataset of AI. The built-in model was then tested using a testing dataset of 20 borderline cases. After running the model on the testing dataset, the accuracy, F1 score, precision and recall were calculated. Results: The present AI model showed an accuracy of 97.97% for extraction and non-extraction decision-making. The receiver operating curve (ROC) and cumulative accuracy profile showed a near-perfect model with precision, recall and F1 values of 0.80, 0.84 and 0.82 for non-extraction decisions and 0.90, 0.87 and 0.88 for extraction decisions. Limitation: As the present study was preliminary in nature, the dataset included was too small and population-specific. Conclusion: The present AI model gave accurate results in decision-making capabilities related to extraction and non-extraction treatment modalities in borderline orthodontic cases of the present population.

show abstract

Section: Methodsmentioning

confidence: 99%

An artificial neural network approach for rational decision-making in borderline orthodontic cases: A preliminary analytical observational in silico study

et al. 2023

View full text Add to dashboard Cite

show abstract

“…A possible improvement of the system is to replace the Raspberry Pi 4B with a more power-efficient SBC, such as a Raspberry Pi Zero 2 W (according to [ 81 ], it has half the power consumption in idle mode), and to use an efficient co-processor for the SDR operations such as a field-programmable gate array (FPGA) [ 82 ], a dedicated digital signal processor (DSP) [ 83 ], a graphics processing unit (GPU) [ 84 ], or tensor processing unit (TPU) [ 85 , 86 ]. The core concept is to leverage the strengths of different architectures for efficient processing [ 87 ]. However, all such approaches require significant hardware and software redesign, which results in higher development costs.…”

Section: Hardware Designmentioning

confidence: 99%

Low-Cost COTS GNSS Interference Monitoring, Detection, and Classification System

Merwe

Franco

Hansen

et al. 2023

Sensors

View full text Add to dashboard Cite

Interference signals cause position errors and outages to global navigation satellite system (GNSS) receivers. However, to solve these problems, the interference source must be detected, classified, its purpose determined, and localized to eliminate it. Several interference monitoring solutions exist, but these are expensive, resulting in fewer nodes that may miss spatially sparse interference signals. This article introduces a low-cost commercial-off-the-shelf (COTS) GNSS interference monitoring, detection, and classification receiver. It employs machine learning (ML) on tailored signal pre-processing of the raw signal samples and GNSS measurements to facilitate a generalized, high-performance architecture that does not require human-in-the-loop (HIL) calibration. Therefore, the low-cost receivers with high performance can justify significantly more receivers being deployed, resulting in a significantly higher probability of intercept (POI). The architecture of the monitoring system is described in detail in this article, including an analysis of the energy consumption and optimization. Controlled interference scenarios demonstrate detection and classification capabilities exceeding conventional approaches. The ML results show that accurate and reliable detection and classification are possible with COTS hardware.

show abstract

“…Several papers have analyzed the cloud version of the TPU. Specifically, [18,19] provide an in-depth survey of the differences and particularities of CPUs, GPUs and (Cloud) TPUs for different tasks related with deep learning. The use of multiple cloud TPUs to scale training tasks has also been deeply studied [11,13].…”

Section: Related Workmentioning

confidence: 99%

Balanced segmentation of CNNs for multi-TPU inference

Villarrubia,

Costero,

Igual

et al. 2023

Preprint

View full text Add to dashboard Cite

In this paper, we propose different alternatives for CNN (Convolutional Neural Networks) segmentation, addressing inference processes on computing architectures composed by multiple Edge TPUs. Specifically, we compare the inference performance for a number of state-of-the-art CNN models taking as a reference inference times on one TPU and a compiler-based pipelined inference implementation as provided by the Google's Edge TPU compiler. Departing from a profiled-based segmentation strategy, we provide further refinements to balance the workload across multiple TPUs, leveraging their co-operative computing power, reducing work imbalance and alleviating the memory access bottleneck due to the limited amount of on-chip memory per TPU. The observed performance results compared with a single TPU yield super-linear speedups and accelerations up to 2.60x compared with the segmentation offered by the compiler targeting multiple TPUs.

show abstract

A Survey of Three Types of Processing Units: CPU, GPU and TPU

Cited by 10 publications

References 3 publications

An artificial neural network approach for rational decision-making in borderline orthodontic cases: A preliminary analytical observational in silico study

An artificial neural network approach for rational decision-making in borderline orthodontic cases: A preliminary analytical observational in silico study

Low-Cost COTS GNSS Interference Monitoring, Detection, and Classification System

Balanced segmentation of CNNs for multi-TPU inference

Contact Info

Product

Resources

About