By: Michael S (already5chosen.delete@this.yahoo.com), May 29, 2022 11:59 am
Room: Moderated Discussions
Linus Torvalds (torvalds.delete@this.linux-foundation.org) on May 29, 2022 10:48 am wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on May 29, 2022 2:15 am wrote:
> >
> > But by now, hopefully, TSX is solidly dead in client CPUs.
> > It should have been like that from the beginning.
>
> Why do you say "in client CPUs"?
Because my name is not Pat Gelsinger.
I would like it to be dead everywhere, but it's not my decision.
And his decision, until at the moment, appears to be "no TSX in the client, but let's drag it for at least one more generation in servers".
>
> Some of the biggest fundamental problems with TSX actually were on the server side. TSX is simply
> an absolutely lovely way to hide your nefarious activity, and do various testing of the speculative
> attacks, and some of them were literally targeted to TSX to begin with (eg TAA).
>
> And the best target for those weren't client CPUs, but cloud.
>
> I really hope that TSX is solidly dead everywhere, because the problem with
> the "don't enable it on client CPUs" is that you then lose developers and test
> cases, and all the problems hit the people who can least afford them.
>
> The whole "this feature is for serious users only" mentality is literally a disease.
> It's how traditional UNIX died, it's how all the workstation/server CPU people died,
> and it's how Intel is going to die if it keeps going down that path.
>
> If something isn't useful in those "client CPUs", then you should
> just say "oh, that's not useful" full stop and just move on.
>
When TSX was first introduced, client CPUs had at most 4 SMP cores. TSX, even if by miracle it works perfectly, simply is not needed with 4 cores. It's simple like that.
And Intel-based servers already had plenty of SMP cores, more than 200 in SGI gear. Here, at least in theory, TSX was potentially useful. Or, at least, it was excusable to believe that it is useful.
> Because mass market is what drives technology. Anybody who loses sight of that has already lost.
>
> Linus
We all know your credo. May be, many even agree. May be, even I agree more than disagree.
But you're, of course, free to repeat it as many times as you want.
> Michael S (already5chosen.delete@this.yahoo.com) on May 29, 2022 2:15 am wrote:
> >
> > But by now, hopefully, TSX is solidly dead in client CPUs.
> > It should have been like that from the beginning.
>
> Why do you say "in client CPUs"?
Because my name is not Pat Gelsinger.
I would like it to be dead everywhere, but it's not my decision.
And his decision, until at the moment, appears to be "no TSX in the client, but let's drag it for at least one more generation in servers".
>
> Some of the biggest fundamental problems with TSX actually were on the server side. TSX is simply
> an absolutely lovely way to hide your nefarious activity, and do various testing of the speculative
> attacks, and some of them were literally targeted to TSX to begin with (eg TAA).
>
> And the best target for those weren't client CPUs, but cloud.
>
> I really hope that TSX is solidly dead everywhere, because the problem with
> the "don't enable it on client CPUs" is that you then lose developers and test
> cases, and all the problems hit the people who can least afford them.
>
> The whole "this feature is for serious users only" mentality is literally a disease.
> It's how traditional UNIX died, it's how all the workstation/server CPU people died,
> and it's how Intel is going to die if it keeps going down that path.
>
> If something isn't useful in those "client CPUs", then you should
> just say "oh, that's not useful" full stop and just move on.
>
When TSX was first introduced, client CPUs had at most 4 SMP cores. TSX, even if by miracle it works perfectly, simply is not needed with 4 cores. It's simple like that.
And Intel-based servers already had plenty of SMP cores, more than 200 in SGI gear. Here, at least in theory, TSX was potentially useful. Or, at least, it was excusable to believe that it is useful.
> Because mass market is what drives technology. Anybody who loses sight of that has already lost.
>
> Linus
We all know your credo. May be, many even agree. May be, even I agree more than disagree.
But you're, of course, free to repeat it as many times as you want.
