System requirements

• At least one graphics card that supports OpenCL 1.2(for future features) (currently uses only OpenCL v1.0 features) and has a dedicated VRAM. Any combo including Nvdia+Amd+Intel should work.

• • In Windows, due to WDDM overhead, Nvidia cards' i/o performance is lowered but with TCC driver mode enabled from nvidia-smi (Some Quadro - Tesla cards have this) it can retain some of performance (currently Ubuntu benchmarks have better results)

• Currently it supports Windows and Ubuntu

• C++14 compiler option enabled (C++1y dialect for g++ compiler)

• for the multi-threaded benchmark in main.cpp, OpenMP is needed ("gomp" library for g++ linker)

• Some RAM that can hold active pages

• • total RAM consumed by active pages = (number of active pages per gpu instance) * (number of gpus) * (4(or custom num with memMult) instances per gpu) * (page size) * sizeof(your_object)
• • if you have 10 gpus, page size=1024, active pages per instance = 100, object size = 100 bytes, then 409MB of RAM will be used
• • • so be careful when using memMult parameter (like {n1,n2,n3,..}), {50,50,50} means 150 gpu instances
• • VRAM usage only changes by number of elements of array (100M elements * 100 bytes per object = 10GB)
• • • Equally distributed between graphics cards
• • • • Or distributed with a ratio as described by memMult parameter: {100,10,1} means first card serves 100x vram, second serves 10x vram, last card serves 1x vram where 111x = total VRAM usage. (same as pcie bandwidth limiting 111x = total pcie bandwidth)