rpe v5: an emulator for reduced floating-point precision in large numerical simulations

Dawson, Andrew; Dueben, Peter

doi:10.5194/gmd-10-2221-2017

Cited by 50 publications

(34 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We also demonstrated that the same basic approach can be applied to a real land surface scheme, using high precision for storing and updating state variables and lower precision for the remaining computations, to achieve a forecast that is within desired accuracy limits. This mixed-precision approach is likely generally applicable to many Earth system model components, and has been demonstrated in other areas (e.g., Düben et al 2017;Dawson and Düben 2017). These results combined suggest that even when it appears lower precision may present problems, it could be possible to continue doing the majority of computations using lower precision and just use higher precision for selected computations that are particularly sensitive to precision.…”

Section: Discussionmentioning

confidence: 72%

Reliable low precision simulations in land surface models

et al. 2017

Self Cite

View full text Add to dashboard Cite

Weather and climate models must continue to increase in both resolution and complexity in order that forecasts become more accurate and reliable. Moving to lower numerical precision may be an essential tool for coping with the demand for ever increasing model complexity in addition to increasing computing resources. However, there have been some concerns in the weather and climate modelling community over the suitability of lower precision for climate models, particularly for representing processes that change very slowly over long time-scales. These processes are difficult to represent using low precision due to time increments being systematically rounded to zero. Idealised simulations are used to demonstrate that a model of deep soil heat diffusion that fails when run in single precision can be modified to work correctly using low precision, by splitting up the model into a small higher precision part and a low precision part. This strategy retains the computational benefits of reduced precision whilst preserving accuracy. This same technique is also applied to a full complexity land surface model, resulting in rounding errors that are significantly smaller than initial condition and parameter uncertainties. Although lower precision will present some problems for the weather and climate modelling community, many of the problems can likely be overcome using a straightforward and physically motivated application of reduced precision.

show abstract

Section: Discussionmentioning

confidence: 72%

Reliable low precision simulations in land surface models

et al. 2017

Self Cite

View full text Add to dashboard Cite

show abstract

“…We use the default of double precision, the new number format, or reduced floating point precision to levels lower than single precision. The latter simulations use the so‐called rpe reduced precision emulator described in Dawson and Düben () to mimic a reduction in floating point precision.…”

Section: Resultsmentioning

confidence: 99%

“…This would reduce the number of bits required for the significand further. However, even with the two changes applied for simulations (as presented in Dawson & Düben, ), differences between the standard model and a model simulation in half precision with 16 bits per variable are visible.…”

Section: Resultsmentioning

confidence: 99%

“…Then, a switch from a standard floating point code to a code that is using the new number format only requires to replace floating point declarations for all numbers that can be different for different ensemble members by the new type format (e.g., replace R E A L ∗ 8:: r 1, r 2, r 3 by TYPE(NEW_FORMAT) :: r 1, r 2, r 3) and to fix model input and model output to be able to cope with the new type and the fact that several ensemble members are calculated within a single simulation. The same application of types in combination with overloaded operators has been used in the past to emulate a reduction of numerical precision beyond single precision (Dawson & Düben, ). Fortran can represent signed integer numbers with 1, 2, 4, and 8 bytes ( n = 8,16,32 or 64 bits).…”

Section: Model and Realization Of The New Number Format In Fortranmentioning

confidence: 99%

See 1 more Smart Citation

A New Number Format for Ensemble Simulations

Dueben

2018

J Adv Model Earth Syst

Self Cite

View full text Add to dashboard Cite

A new number format for ensemble simulations for weather and climate predictions is presented. The new format exploits similarities between ensemble members to reduce overall data usage. Data use and movement is the most important performance bottleneck for weather and climate models on modern supercomputers. The new format can be realized on standard hardware and with standard computing languages such as Fortran. The format is tested in simulations with a shallow water model. Advantages and disadvantages are discussed.

show abstract

“…To further verify our methods, simulations were also conducted with 16-bit floating-point numbers, which is referred to as half precision by the IEEE 754 standard. Because precisions lower than 32 bits are generally unavailable from the computer systems and Fortran compilers EAMv1 is supported for, the reducedprecision emulator from Dawson and Düben (2017) was used. For each emulated half-precision number, 5 of the 16 bits were used for the exponent and 1 for the sign, leaving 10 bits for the mantissa, hence the half-precision number only had three to four accurate decimal digits.…”

Section: Introducing Reduced Precision To Eamv1mentioning

confidence: 99%

An Objective and Efficient Method for Assessing the Impact of Reduced‐Precision Calculations On Solution Correctness

Zhang

Wan

Rasch

et al. 2019

J Adv Model Earth Syst

View full text Add to dashboard Cite

Recent studies have shown that reducing the precision of floating‐point calculations in an atmospheric model can improve the model's computational performance without affecting model fidelity, but code changes are needed to accommodate lower precision or to prevent undue round‐off error. For complex and computationally expensive systems like the Energy Exascale Earth System Model, a method is needed to objectively and efficiently assess the quality of the lower‐precision simulations and to quickly identify problematic code pieces. This paper demonstrates that the accuracy of the computed solution can be evaluated through a simple and quantitative error metric based on time step convergence. The proposed test method can unambiguously detect the accumulation of precision error and objectively assess its impact on solution accuracy. Furthermore, since fast physical processes are known to affect key features of the multiyear mean climate in atmospheric models, we show that the convergence test applied to short simulations can provide useful information about the impact of reduced precision on a model's long‐term behavior. In contrast, the traditional method of climate model evaluation requires multiple years of simulations and involves inspecting many physical quantities for statistical significance in the presence of natural variability. The simplicity, computational efficiency, and objectivity of the proposed method makes it very attractive for high‐resolution model development. The method is also expected to be applicable to other models that numerically solve time evolution equations.

show abstract

rpe v5: an emulator for reduced floating-point precision in large numerical simulations

Cited by 50 publications

References 25 publications

Reliable low precision simulations in land surface models

Reliable low precision simulations in land surface models

A New Number Format for Ensemble Simulations

An Objective and Efficient Method for Assessing the Impact of Reduced‐Precision Calculations On Solution Correctness

Contact Info

Product

Resources

About