< Previous | Contents | Manuals Home | Boris FX | Next >
GPU Preferences
These controls are specific to your machine, and govern how or if Tensorflow or macOS attempts to use the GPU on your machine. Note that you must have a suitable GPU and the Windows or Linux Tensorflow-GPU overlay installed in your system, see the Installing the Tensorflow-for-GPUs Overlay section above. (Again, macOS 10.15+ already has the necessary support.)
GPU Control. Selector between Auto, CPU only, or “GPU, or CPU”. In Auto mode, Tensorflow/macOS gets to decide. In CPU only mode, Tensorflow/macOS is told NOT to use the GPU. In “GPU, or CPU” mode, we tell Tensorflow to use one, but it is still at its discretion whether or not it is able to do so. While the behavior of “GPU, or CPU” and Auto should be the same, we can’t guarantee that Tensorflow treats them as the same. Important: on macOS systems with ANE, the neural engine is (possibly) enabled only if Auto is selected, the “GPU, or CPU” mode specifically excludes the ANE. In all cases, there is no way to force processing onto an unwilling GPU or ANE.
Use minimum GPU RAM. Checkbox. Normally, Tensorflow allocates and retains ALL memory on the GPU (that isn’t already used for monitor display). Obviously, that isn’t too friendly. The Use minimum GPU RAM tells Tensorflow to only ask for memory
when it needs it… but it is really hungry, so this probably won’t matter unless you have tens of GBs of RAM on your GPU, or keep the tile settings small.
GPU memory %. Spinner. When non-zero, we tell Tensorflow to use only this much of the GPU’s memory…maybe 25% or 50% would be a typical value, if the card is to do something else big. The downside of using this spinner is that by limiting the memory available, you’ll have to use smaller tile settings to stay within that available memory.
©2024 Boris FX, Inc. — UNOFFICIAL — Converted from original PDF.