RMM
23.12
RAPIDS Memory Manager
|
Achieving optimal performance in GPU-centric workflows frequently requires customizing how host and device memory are allocated. For example, using "pinned" host memory for asynchronous host <-> device memory transfers, or using a device memory pool sub-allocator to reduce the cost of dynamic device memory allocation.
The goal of the RAPIDS Memory Manager (RMM) is to provide:
For more information on APIs provided by rmm, see the modules page.