Some AMD Ryzen CPUs don't run as hot as is reported

[DanceswithUnix]

Still think a much-needed feature in modern OSes is a "disable SMT for this program" option to have the OS report an 8C8T processor to the software and have the scheduler lock the affinity of that software to primary cores only. Even with perfect SMT there is plenty of software that is disadvantaged by it, but it is a bit extreme to disable it globally.

Perhaps software should have a thread profile for task schedulers to read and better assign threads as well, e.g. to assign heavy single threaded software to a core whilst keeping the associated SMT core clear of tasks.

It would be really easy to implement a "soft" SMT disable, just get the sheduler to not allocate work to the second thread of each core. Job done.

Sadly I don't think the world is simple enough for that to be useful. What if a program has enough threads that you want SMT turned off with an 8 core CPU but you want it on if you only have 2 or 4 cores? What if you are playing a game that turns off SMT in your 8 core, but you want to live stream your game which really requires another 4 threads from the remaining 2 cores where unfortunately you turned SMT off? So now you want "for the cores used by this game don't use SMT, but you can use it on other cores used by other programs". Starting to get complex now, how does the OS decide which cores to use for which programs? How many "SMT off" programs do you have running before you realise there are so many threads that you would actually be better off ignoring the hint and turning SMT on?

Scheduling is hard, I suspect the answer here is some evolutional or deep learning system to tweak things until optimal. Too many variables for a human.

[kompukare]

Just build normal multi-purpose CPUs like Intel makes you'll be successful in the market.

Ah, it's funny you should mention Intel in your rant.
Tell, what is an Intel Core i7? To make it somewhat easier we'll ignore HEDT here, so is it:

1. 4C/8T ?
2. 2C/4T?

And what is an Intel Core i5?

1. 4C/4T?
2. 2C/4T?

Well, actually in both cases the answer is both as mobile i7s and i5s are essentially similar to desktop i3s. Well no, that's not quite true as Intel do make some mobile i7s which are 4C/8T. And recently some mobile i5s which are 4C/4T.

Confused yet? How about Xeon E3? Are they all at least 4C/8T as you would expect? No, they also make 4C/4T ones. Virtualisation features? AVX2 etc?

No when it comes to confusing the consumer, Intel are experts. In fact, not even the various helpful Wikipedia listings of their processor (like the Xeon list, or the i7 list) are not enough so you have to go ark.intel.com to be sure.

[scaryjim]

... now you want "for the cores used by this game don't use SMT, but you can use it on other cores used by other programs". ...

I suspect some games attempt to imitate this behaviour by querying Windows for the CPU core configuration and then setting core affinity masks programmatically. Nice idea in principle, but if a new CPU comes along and your game doesn't understand its core config properly that could cause problems.

As could Windows 10 not handling core affinity correctly, which at least one or two testers have demonstrated on Ryzen...

[DanceswithUnix]

Nice idea in principle, but if a new CPU comes along and your game doesn't understand its core config properly that could cause problems.

If they do that then they should really hire better programmers. If the code comes across a config it doesn't recognise, then in the context of a game it should just leave the heck alone, leaving a debug trace in a log file somewhere to say it needs fixing. You don't guess.

[kalniel]

As could Windows 10 not handling core affinity correctly, which at least one or two testers have demonstrated on Ryzen...

Have you got links to that? I thought it'd been quite comprehensively proved Windows 10 is handling core affinity correctly- both avoiding SMT 'cores' and keeping to the same CCX where possible.
https://www.pcper.com/reviews/Proces...-Silver-Bullet

[scaryjim]

It was in the Strictly Technical anandtech forum thread - one of the users enabled SMT, switched to a 4+0 core setup and masked a heavily threaded process to 0,2,4,6. The scheduler only allocated to cores 0 & 2 which were maxed out. I'll stick a link up when I find it again...!

EDIT: found it; this post from looncraz: https://forums.anandtech.com/threads...#post-38789200

Sets cinebench to 8 threads, masks to 4 cores, finds only 2 get loaded while the other two are parked ... in high performance mode.

If that's happening in games that set their own affinity masks (and I suspect quite a few do in order to "fix" perceived Windows Scheduler issues), it could effectively make an 1800X perform like a dual core CPU...

[kalniel]

So in that case, it's actually sticking to within the same CCX rather too well. Interesting.

[scaryjim]

So in that case, it's actually sticking to within the same CCX rather too well. Interesting.

It's more than that - all 4 cores are in the same CCX (it's a 4+0 downcore), so it's just plain wrong. For some reason it's misreading the affinity mask as just 0,2. Particularly curious is that he says it doesn't happen for a 0,1,2,3 mask, and it doesn't happen when SMT is disabled. So something curious is definitely up.

Just been browsing the rest of that thread. Looncraz has put a lot of effort into testing (on Win 7 as well as Win 10) and comes up with some very interesting conclusions, most of which support everything AMD's been saying so far...

[kalniel]

Oh I see, weird.

[scaryjim]

Oh I see, weird.

I think that might be an understatement

Definitely something odd going on. It's going to be very interesting to see how AMD tweak the Zen design over the next couple of revisions. General consensus is that if they can speed the fabric up that'll mitigate the vast majority of cases where the CCX design is causing issues. Amidst all the discussion of there being a hidden multiplier for the fabric that would run it at twice the memory clock I can't help wondering if GF's 14nm process simply can't sustain the fabric at that high a clock rate within the voltages/heat dissipation targets that AMD set for Ryzen....

I also wonder if most games wouldn't run better if they just let the Windows scheduler handle threading for them instead of trying to "optimise" it directly...

[kalniel]

I also wonder if most games wouldn't run better if they just let the Windows scheduler handle threading for them instead of trying to "optimise" it directly...

And we wonder why many game devs code largely for single threaded performance

[scaryjim]

And we wonder why many game devs code largely for single threaded performance

I remember messing around with threading code while I was doing my MSc. I don't wonder that at all...

[scaryjim]

Another result has been posted in the Strictly Technical thread showing yet another instance of core parking/balanced power mode messing up a benchmark result:

https://forums.anandtech.com/threads...#post-38799170

Testing winRar with affinity to bind to a CCX: on cores 0-7 runs fine, on cores 8-15 performance tanks. The second result was only using 2 threads, with 3 cores (6 threads) parked. Apparently switching to Performance power mode fixed the problem.

Definitely issues with power plans, core parking and core affinity between Win 10 and Ryzen....