Linux kernel boost
Heterogenous memory management lends a helping hand.
Heterogenous memory management (HMM) could soon be making its way to the Linux kernel, which could give the OS a major boost when it comes to machine-learning applications and other GPU-powered tasks.
HMM enables device drivers to mirror the address space for a process under its own memory management. So a graphics card could directly access the memory of a process without having to copy anything or putting at risk an OS’s memory protection feature.
Basically, this can give a hefty speed boost to GPU-based machine learning, with libraries such as Nvidia’s CUDA, as well as OpenCL. While this is an exciting development, you’ll need to keep your expectations in check for now. For a start, HMM needs to be committed to the Linux kernel, and this was first proposed back in 2014, so it obviously is taking quite a while to implement. It’s also likely that you will need a new – and expensive – graphics card to make use of HMM – at the moment only Nvidia’s high-end Pascal GPUs support the feature. So, while there may be a bit of time left until we see HMM support in the kernel, we’d best start saving for a compatible GPU now.