Back to Table of contents

Primeur weekly 2018-03-19

Exascale supercomputing

BSC and Fujitsu to build up an ecosystem for a future exascale supercomputer ...

Quantum computing

New quantum spin liquid predicted by Nobel Laureate prepared for the first time ...

Focus on Europe

New project to plug the software-stack support gap for energy-efficient computing ...

North German scientists receive new supercomputer in the HLRN network ...

Hardware

Cray tapped to deliver largest supercomputer dedicated to fusion science in Japan by National Institutes for Quantum and Radiological Science and Technology ...

Keeping GPUs young ...

Asetek announces ongoing collaboration with Intel on liquid cooling for servers and data centres ...

UC San Diego Center for Microbiome Innovation announces Panasas as Corporate Member ...

Cavium expands industry partnerships to drive OCP NIC 3.0 ...

Applications

How much snow accumulates in North America each year? More than scientists thought. ...

Green light for Galaxy Europe ...

Supercomputer simulation opens prospects for obtaining ultra-dense electron-positron plasmas ...

The complex journey of red bloods cells through microvascular networks ...

The view from inside supersonic combustion ...

A game changer: Metagenomic clustering powered by supercomputers ...

ORNL researchers design novel method for energy-efficient deep neural networks ...

University of Texas researchers develop new visualization tools to explore fusion physics ...

Fishing with simulations for two-dimensional materials ...

The Cloud

IBM helps accelerate AI with fast new data platform and elite team ...

European Commission to adopt Implementation Roadmap for the European Open Science Cloud ...

Univa announces a global partnership and reseller agreement with UberCloud ...

Univa open sources Project Tortuga to accelerate the migration of enterprise HPC workloads to the Cloud ...

USFlash

Xilinx unveils revolutionary adaptable computing product category ...

Keeping GPUs young

12 Mar 2018 Vienna - Graphics processing units (GPUs) are not only used for displaying graphics. Today, they are frequently used for particularly challenging calculations - for example in scientific research or even Bitcoin-mining. Graphics processing units are used for many computationally intensive tasks. Their aging process can be slowed by clever management, as TU Wien and University of California - Irvine have now shown.

However, their performance usually decreases over time. As the individual cores of the processing unit age, they do not work together perfectly any more. TU Wien and the University of California, Irvine have therefore developed an improved chip management method, which distributes the processing tasks in an efficient way. In more than 95% of cases, this can slow down the aging process of GPUs.

"The standard computers that we are using at home just have a few processor cores. A graphics processing unit on the other hand consists of a large number of cores - typically hundreds or thousands", stated Professor Muhammad Shafique, Institute for Computer Engineering, TU Wien. Therefore, GPUs are extremely useful for parallel calculations, in which a task can be broken down into many smaller calculations that can be solved independently from each other, by different cores at the same time.

The cores, however, are never completely identical. "The structures on the GPU are very small. There will always be small imperfections and deviations due to the production process", stated Muhammad Shafique. Therefore, aging has different effects on different cores. Under stress, the performance of different cores will drift apart.

This is a problem for parallel computing. When a task is distributed among many cores and the system has to wait for the last one to finish before the next step of the calculation can be started, the slowest core determines the overall speed of the process. This can drastically reduce the performance of the GPU.

Muhammad Shafique and the team of Professor Mohammad Al Faruque, University of California, Irvine want to solve this problem using clever management methods: "First, it is crucial to know the status of the individual cores", stated Muahmmad Shafique. "Then we can distribute different tasks in such a way among the cores that physical stress is optimally distributed."

Cores with similar properties are combined to "clusters", which are then assigned with more or less challenging tasks - depending on their current status. That way, the aging process can be slowed down, the overall performance of the system increases. "There have already been ideas on how to slow the aging process of GPUs with compiler-based techniques, but our method is much more effective. It works in more than 95% of cases", stated Muhammad Shafique.

The paper titled " Aging-aware Workload Management on Embedded GPU Under Process Variation " published in the journalIEEE Transactions on Computershas been featured byNature Electronicsas a "Research Highlight": " Just about managing ".

Source: Vienna University of Technology

Back to Table of contents

Primeur weekly 2018-03-19

Exascale supercomputing

BSC and Fujitsu to build up an ecosystem for a future exascale supercomputer ...

Quantum computing

New quantum spin liquid predicted by Nobel Laureate prepared for the first time ...

Focus on Europe

New project to plug the software-stack support gap for energy-efficient computing ...

North German scientists receive new supercomputer in the HLRN network ...

Hardware

Cray tapped to deliver largest supercomputer dedicated to fusion science in Japan by National Institutes for Quantum and Radiological Science and Technology ...

Keeping GPUs young ...

Asetek announces ongoing collaboration with Intel on liquid cooling for servers and data centres ...

UC San Diego Center for Microbiome Innovation announces Panasas as Corporate Member ...

Cavium expands industry partnerships to drive OCP NIC 3.0 ...

Applications

How much snow accumulates in North America each year? More than scientists thought. ...

Green light for Galaxy Europe ...

Supercomputer simulation opens prospects for obtaining ultra-dense electron-positron plasmas ...

The complex journey of red bloods cells through microvascular networks ...

The view from inside supersonic combustion ...

A game changer: Metagenomic clustering powered by supercomputers ...

ORNL researchers design novel method for energy-efficient deep neural networks ...

University of Texas researchers develop new visualization tools to explore fusion physics ...

Fishing with simulations for two-dimensional materials ...

The Cloud

IBM helps accelerate AI with fast new data platform and elite team ...

European Commission to adopt Implementation Roadmap for the European Open Science Cloud ...

Univa announces a global partnership and reseller agreement with UberCloud ...

Univa open sources Project Tortuga to accelerate the migration of enterprise HPC workloads to the Cloud ...

USFlash

Xilinx unveils revolutionary adaptable computing product category ...