- 积分
- 57
- 注册时间
- 2006-6-24
- 仿真币
-
- 最后登录
- 1970-1-1
|
Graphics Processing Units (GPUs) can be used to accelerate the Algebraic Multigrid (AMG) solver inside
ANSYS FLUENT. GPU acceleration is currently only available for the pressure coupled equation system
arising from the 3D pressure-based coupled solver.
You can enable GPU acceleration using the solve/set/multi-grid-amg text user interface (TUI)
command. An NVIDIA Fermi or later GPU is required along with CUDA 4.1. GPU acceleration of the AMG
solver is limited to serial and shared memory parallel FLUENT sessions using a single GPU on lnamd64
(Red Hat Enterprise Linux 5/6, and SUSE Linux Enterprise Server 11) and win64 (Windows 7) platforms.
As with any CPU system, the problem size that can be solved using a single GPU will depend on the
amount of memory available on the GPU. The AMG cycle-type, coarse group size, and smoother weights
all can impact convergence and performance. For instance, the V-cycle type converges more quickly
than the default F-cycle type on the GPU. The group sizes for GPU execution are limited to 2, 4, and 8.
Using a larger group size is faster, however, using a smaller group size leads to improved convergence.
In addition, a smoother relaxation/weight can improve convergence for complicated problems (e.g.,
you can use the Scheme command (rpsetvar 'amg/ilu-cpld-relax 0.5) in order to explicitly
set the relaxation to 0.5).
|
|