high performance computing on graphics processing units: hgpu.org

This task involves the following work: – Establishing the relation of Quality of Service (QoS) and energy to accuracy. – Design and development of techniques to dynamically decrease accuracy (e.g., ignore low order bits in computations). Deliberately ignoring a few low order bits in calculations where the application allows it (in terms of impact to QoS) reduces energy requirements. We will implement techniques to allow change of accuracy, dynamically (run-time) or under application control. – Design and development of techniques to tolerate transient errors in computations. Reducing the supply voltage beyond the point required by the operational frequency can yield significant power reductions but at the expense of potential (timing) errors. Such errors sacrifice some of the accuracy of computations. However, if such errors are not important to the resulting quality of service then this is a promising approach to increase power efficiency. In cases where absence of any error is necessary for the correct operation of the GPU, error detection and correction techniques will be applied.