|
Titel |
Inexact hardware and the trade between precision and performance in earth system modelling |
VerfasserIn |
Peter D. Düben, Stephen Jeffress, Tim N. Palmer |
Konferenz |
EGU General Assembly 2015
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 17 (2015) |
Datensatznummer |
250106746
|
Publikation (Nr.) |
EGU/EGU2015-6426.pdf |
|
|
|
Zusammenfassung |
We study the use of inexact hardware in numerical weather and climate models. Inexact
hardware is promising a reduction of computational cost and power consumption of
supercomputers and could be a shortcut to higher resolution forecasts with higher forecast
accuracy and exa-scale supercomputing. However, simulations with inexact hardware show
numerical errors, such as rounding errors or bit flips.
In cooperations with groups in computing science, we studied different approaches to
inexact hardware that include the use of stochastic processors: the applied voltage in
computing hardware is reduced to save power, but bit flips are possible, the use of
pruned hardware: parts of the floating-point unit that are either hardly used or do not
influence significant bits are removed, the use of Field Programmable Gate Arrays
(FPGAs): An FPGA is a programmable hardware that allows flexible floating-point
precision, and the use of inexact memory within simulations of numerical models
for weather and climate predictions. Results show that numerical precision can be
reduced significantly within simulations of the three dimensional atmosphere with no
significant increase in model errors. If computational cost is reduced due to the
use of inexact hardware, the possible increase in resolution will allow a stronger
reduction of model errors compared to the increase of model errors due to reduced
precision.
We treat different parts of atmospheric models with customized computational accuracy to
reflect inherent uncertainties. Planetary scale waves are more predictable and less uncertain
than meso-scale waves. For small-scale dynamics, diffusion, parametrisation schemes, and
sub-grid-scale variability cause large inherent uncertainties. An approach of scale separation
that calculates the dynamics of expensive small scales with low numerical precision and
the dynamics of large scales with high precision has proved to be very efficient. |
|
|
|
|
|