By: Dummond D. Slow (mental.delete@this.protozoa.us), November 20, 2020 9:45 am
Room: Moderated Discussions
Ungo (a.delete@this.b.c.d.e) on November 19, 2020 8:12 pm wrote:
> Dummond D. Slow (mental.delete@this.protozoa.us) on November 19, 2020 3:03 pm wrote:
> > What I think is likely:
> > 1) SMT support needs less area than a small core (unless it is something like small Cortex-M obviously).
> > 2) I am less certain but still tilting that it also might
> > cost less power. Amongst other things because their
> > improve utilization of existing units instead of needin gnew units that would be poweres in parallel to new
> > ones. Not a given, some small cores might be frugal enough (but then would they have enough performance)?
> > 3) SMT definitely makes core harder to design and validate.
> > 4) I think the side-channel risk a wrong thing to look at so I ignore that.
> >
>
> On point 2 - I think you have missed the primary reason to have a small core. It is to support ultra
> low power states when the system is lightly loaded and none of the software running requires high performance.
> Scenarios like that are very common in battery powered personal computing devices, and are the true
> point of the little cores. If the only things you're doing with a CPU for ten seconds are redrawing
> the mouse cursor and pinging the user's IMAP server to see if there are new emails, it's far more energy
> efficient to do that work with a little core while all the big cores are power-gated.
>
> It seems plausible to me that you are correct and SMT might be a path to better perf/W when the system
> is under heavy load, but little cores are really about other scenarios. Letting them provide some
> extra throughput during heavy load scenarios makes sense, if they have to be there anyways.
>
> Also, I can't disagree more that side-channel risk is a wrong thing to look at.
Yes but SMT is not the only side channel opportunity, removing it only "saves" you from one avenue of great many, the rest of which will remain = you gain close to nothing. Hence it's silly to oppose it on that reasoning, guarding against side channels is really going to be mainly OS's responsibility, I don't think we'll eradicate all of them from hardware.
> Dummond D. Slow (mental.delete@this.protozoa.us) on November 19, 2020 3:03 pm wrote:
> > What I think is likely:
> > 1) SMT support needs less area than a small core (unless it is something like small Cortex-M obviously).
> > 2) I am less certain but still tilting that it also might
> > cost less power. Amongst other things because their
> > improve utilization of existing units instead of needin gnew units that would be poweres in parallel to new
> > ones. Not a given, some small cores might be frugal enough (but then would they have enough performance)?
> > 3) SMT definitely makes core harder to design and validate.
> > 4) I think the side-channel risk a wrong thing to look at so I ignore that.
> >
>
> On point 2 - I think you have missed the primary reason to have a small core. It is to support ultra
> low power states when the system is lightly loaded and none of the software running requires high performance.
> Scenarios like that are very common in battery powered personal computing devices, and are the true
> point of the little cores. If the only things you're doing with a CPU for ten seconds are redrawing
> the mouse cursor and pinging the user's IMAP server to see if there are new emails, it's far more energy
> efficient to do that work with a little core while all the big cores are power-gated.
>
> It seems plausible to me that you are correct and SMT might be a path to better perf/W when the system
> is under heavy load, but little cores are really about other scenarios. Letting them provide some
> extra throughput during heavy load scenarios makes sense, if they have to be there anyways.
>
> Also, I can't disagree more that side-channel risk is a wrong thing to look at.
Yes but SMT is not the only side channel opportunity, removing it only "saves" you from one avenue of great many, the rest of which will remain = you gain close to nothing. Hence it's silly to oppose it on that reasoning, guarding against side channels is really going to be mainly OS's responsibility, I don't think we'll eradicate all of them from hardware.