Skylake: SVM / Shared Virtual Memory

celli4 · ‎10-17-2015

Hello,

I would like to ask about the Shared Virtual Memory (SVM) feature of Skylake as it relates to Quick Sync.

Currently on Haswell, and older, even though video memory and general cpu memory come off of the same physical chips or DIMMs, they are partitioned effectively from the software developers perspective:

ie, When you decode H.246, you are decoding to system memory or to video memory. They are effectively two memory types.

[actually all decodes are done to video memory, but the IMSDK will internally copy from vid-mem to system memory when a decoder is setup for system memory]

And moving frames/surfaces between the two areas is not cheap performance-wise.

Will the GPU SVM feature of SKL change how things operate somehow? Please explain where we can learn more about how the Shared Virtual Memory (SVM) feature of SKL relateds to Quick Sync surfaces and how things have changed from Haswell.

Thank you, Cameron

Jiandong_Z_Intel · ‎10-19-2015

Hi Cameron,

There's a whitepaper on Intel’s website detailing the compute architecture of Intel's "Gen9" graphics. You can search with key word “SVM”,

“Coherent SVM write performance is significantly improved via new LLC cache management policies”

Graphic driver will use HW features.

Hope this can help.

Thanks,

Zachary