8.7. Platforms
This release of OpenMM contains the following Platform subclasses:
ReferencePlatform: This is designed to serve as reference code for writing other platforms. It is written with simplicity and clarity in mind, not performance.
CpuPlatform: This platform provides high performance when running on conventional CPUs.
CudaPlatform: This platform is implemented using the CUDA language, and performs calculations on Nvidia GPUs.
OpenCLPlatform: This platform is implemented using the OpenCL language, and performs calculations on a variety of types of GPUs and CPUs.
The choice of which platform to use for a simulation depends on various factors:
- The Reference platform is much slower than the others, and therefore is rarely used for production simulations.
- The CPU platform is usually the fastest choice when a fast GPU is not available. However, it requires the CPU to support SSE 4.1. That includes most CPUs made in the last several years, but this platform may not be available on some older computers. Also, for simulations that use certain features (primarily the various “custom” force classes), it may be faster to use the OpenCL platform running on the CPU.
- The CUDA platform can only be used with NVIDIA GPUs. For using an AMD or Intel GPU, use the OpenCL platform.
- When running on recent NVIDIA GPUs (Fermi and Kepler generations), the CUDA platform is usually faster and should be used. On older GPUs, the OpenCL platform is likely to be faster. Also, some very old GPUs (GeForce 8000 and 9000 series) are only supported by the OpenCL platform, not by the CUDA platform.
- The AMOEBA force field only works with the CUDA platform, not with the OpenCL platform. It also works with the Reference and CPU platforms, but the performance is usually too slow to be useful on those platforms.
.gif "sm_B-) :-)")
