10-18-2018, 06:06 AM
(10-17-2018, 08:59 AM)philsmd Wrote: As epixoip mentioned there is a limit for several OpenCL drivers that each memory allocation can't be more than 1/4 of the available VRAM. if you, in theory, allocate 4 times (multiple times) you have 4 * 1/4 = 1 and therefore could in theory use the full VRAM available.... The problem is that there is extra cost (and speed drop) in maintaining more than one separate buffers/allocations, so that is not really a good solution except for a few exceptions like scrypt etc (where we have some work around in place already)Thanks for explaining.
The benefit of large GPU memory buffers is of course to load more digests, have more room for longer rules (with more rule functions), allow different TMTO settings for algos like scrypt etc.
The problem often is that the GPU has e.g. 8 GB ram, but you can only allocate 1/4 * 8 = 2 GB of 8 GB (with one single allocation/buffer) because of that silly restriction (miss interpretation of the spec).