BEGIN:VCALENDAR
PRODID:-//Columba Systems Ltd//NONSGML CPNG/SpringViewer/ICal Output/3.3-
M3//EN
VERSION:2.0
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
DTSTAMP:20191115T102823Z
DTSTART:20191129T140000Z
DTEND:20191129T150000Z
SUMMARY:Jan Ackmann - Reduced Precision Computing for Weather and Climate
Models
UID:{http://www.columbasystems.com/customers/uom/gpp/eventid/}pw0-k20675p
o-x39knl
DESCRIPTION:Jan Ackmann from University of Oxford will be speaking at the
seminar.\n\nAbstract :\nWeather and Climate (W&C) prediction models are
required to satisfy strict time-to-solution and energy-to-solution cons
traints and thus need to be as computationally efficient as possible on
modern supercomputers. To explore possible gains in computational effici
ency\, we follow various approaches: Precision reduction for floating po
int operations\, exploring alternative number formats\, and replacing mo
del components with machine-learned surrogates. Resulting computational
savings could then be reinvested where they are needed more urgently.The
justification for these approaches – that inevitably introduce addition
al model errors – is motivated by the presence of irreducible uncertaint
ies in W&C model predictions. These uncertainties are due to the interpl
ay of the W&C models’ two main components\, the dynamical core\, a discr
etization of the Navier-Stokes equations\, and the so-called model physi
cs – a collection of stochastic parametrizations for the unresolved subg
rid-scale processes (turbulence\, cloud physics\, convection\,…). In the
presence of the resulting uncertainties\, high precision is deemed unne
cessary for many computational operations and model components.\nThe fir
st part of the talk will be about the group’s work on the use of low pre
cision arithmetic for various model components (Spectral dynamical cores
\, Adjoint calculation\, Data Assimilation\, Legendre Transforms\, and m
odel physics)\, where often a level of half precision is found feasible.
Also\, alternative number formats such as Posits (emulated in software)
and neural network approaches that replace parts of the model physics a
re discussed.\n\nThe second part of the talk will be a more detailed acc
ount on reduced-precision preconditioned elliptic solvers in dynamical c
ores.
STATUS:TENTATIVE
TRANSP:TRANSPARENT
CLASS:PUBLIC
LOCATION:Frank Adams 1\, Alan Turing Building\, Manchester
END:VEVENT
END:VCALENDAR