What impact on DOTS does the rising trend in big.LITTLE core CPU architectures have?

Mobile CPUs have had big cores and little cores partly from the transition between 32 and 64 bit features and more recently in large more mainstream CPUs e.g. the new Apple M1 and it sounds like next gen AMD chips will follow this trend.

So what impact on DOTS does this trend in big.LITTLE cores have on performance, if the very hardware constrains some threads and not others.

If DOTS is taking advantage of the hardware do we need big and LITTLE systems in DOTS?

Does this limit any approach that would allow interleaving DOTS systems to fully utilise the hardware e.g. a big core could have way more L0 cache/registers and SIMD power than a LITTLE core?

 

I wonder if Unity will provide a way to choose which type of core you can execute your jobs on?

 

Wouldn't you pick one type of target and optimize for that? Aren't you either looking to make an app that is low power or high performance, not both? If you want to be able to perform well on low power cores then it seems a high power core would handle you're app well enough anyway, optimized or not. If on the other hand you need the performance of a high power core then you've already eliminated the low power core as an option.

 

You could have a single threaded job (chain) that you know will be part of the main execution path; having that run on the high power cores while the "other" jobs not directly on the main path run on the rest would probably be desirable.

 

DOTS regarding different architectures and cores / affinities seems like something we could control in a few years. The core idea of DOTS is going to be the same for every problem here but the tweaking is not, and I don't expect the magic numbers to change much while in dev.

By that I mean the general resource behaviour and distribution/size of work.

 

DOTS also seems to improve performance so much due to marshaling of resources correctly that the little cores and the big cores may have negligible difference in absolute performance. Or I could be wrong.

 

I suspect that the OS will ultimately decide what work to put on which type of cores. A game's threads will probably be run on high performance cores while background tasks will be put on the efficiency cores.

 

Interesting article on how Apple runs the OS on the LITTLE cores and reserves the four big cores for user apps/tasks Clever OS Scheduling Partly Explains Apple M1's Responsiveness - ExtremeTech.

So in theory any game on Apple M1 should have four big cores to run on at least.