Single Precision Computation

The single precision version of Radioss is referred to as extended single precision, since most calculations are single precision but some remain double precision to obtain a more accurate result.

By default, Radioss uses a double precision floating-point format when completing a calculation as opposed to other explicit solvers which utilize single precision by default. The double precision format uses from 15 to 17 significant decimal digits precision; whereas, the single precision format uses 6 to 9 significant decimal digits precision.

Although model dependent, the single precision version runs 30% - 40% faster than the double precision version.

When single precision is used in a calculation, there is more round off error in the floating point calculations due to the reduced number of significant digits in single precision. Since explicit calculations calculate one cycle at a time and the next cycle is based on the previous cycles calculations, simulations with a large number of cycles can accumulate rounding error as the simulation progresses resulting in a less precise result.

To minimize round off error, it is recommended to use the default version of Radioss which is double precision. This is most important in simulations that have these characteristics:
  • Larger number of solution cycles due to a small time step or long simulation time.
  • Small element size
  • Airbag simulations
  • Implicit simulations
  • Fluid and Fluid - Structure interaction simulations, ALE, CEL
  • High-speed impact such as ballistic simulations

The easiest way to determine if a simulation will give accurate results using single precision calculations is to compare the single precision results to the double precision results to identify any large differences. The single and double precision results will not be identical, but they should be very close.

Start a Single Precision Simulation

When using the HyperWorks Solver Run Manager, the single precision version can be launched using the -sp option.

radioss -np 24 -sp tensile_model_0000.rad

If using the Radioss executables directly, the single precision version is ran using the executables that include the “_sp” name for both the Radioss Starter and Engine.

./s_<version>_<platform>_sp –i <prefix>_0000.rad

./e_<version>_<platform>_sp –i <prefix>_0001.rad