Windows 10 Scheduler bug causes poor RyZen performance

[Space Reptile]

5 hours ago, LAwLz said:

OK

So according to you the vast majority of games will see a ~20% increase in FPS if the scheduler gets patched?

I think you are out of your mind and is setting yourself up for a massive appointment, but we will see.

i never claimed 20% , again read what people post, you get about 10% to AT BEST 20% 

 

but who am i kidding ,

AMD is confirmed dead and i cant wait for the 700$ 4c/8t i7 thanks to monopoly ,

because it has to be good when its expensive 

 

OBVIOUS FUCKING SARCASM FOR THE DEAF AND BLIND 

Edited March 9, 2017 by wkdpaul
cleaned up

[Misanthrope]

Bugs to neutralize eh?

 

Why does it works normally on Windows 7 which according to you shouldn't even work at all Mr. Bullshit Artist Microsoft Employee?

[LAwLz]

5 hours ago, Space Reptile said:

i never claimed 20% , again read what people post, you get about 10% to AT BEST 20% 

Ehm...  The conversation went like this:

 

Person: You get up to 20% performance improvement with SMT disabled so it won't be lower than that.

Me: Do you honestly think that we will get a 20% performance improvement?

Person: It will happen in some games!

Me: Got any examples of games where this will happen?

You: Games that uses less than 6 cores/8 threads will see this improvement.

Me: So you think most games will get a 20% performance increase?

You: I never claimed that! Use your brain! They will get between 10-20%!

 

If you follow the reply chain you will see that this is how the conversation went...

 

 

5 hours ago, Space Reptile said:

but who am i kidding ,

AMD is confirmed dead and i cant wait for the 700$ 4c/8t i7 thanks to monopoly ,

because it has to be good when its expensive 

 

OBVIOUS FUCKING SARCASM FOR THE DEAF AND BLIND 

Is that suppose to be me? Because I've been saying this since Ryzen reviews came out:

LGA 1151 - Still has no competition. If you were planing on getting one, then get one.

LGA 2011-3 - Dead. Ryzen killed it except for the very, very niche market (and things like servers, obviously).

Ryzen 7 - What you should get if you were thinking of getting something like the 6800K.

 

The only thing I have declared dead is X99, for most consumers at least.

 

I get that you want AMD to succeed and is scared of an Intel monopoly, but that's not excuse for buying into the hype machine or giving people hope that is just based on quite frankly ridiculous speculations.

 

Some first line support person from Microsoft tweeted that they will look into Ryzen? OMG Windows 10 scheduler confirmed for crippling Ryzen! Up to 20% gaming performance increase confirmed guys!

 

 

The reason why I bring up Bulldozer is because people said the exact same thing back then. They were always blaming something other than AMD and they kept getting disappointed over and over again. Lower your expectations and look at how Ryzen performs today, not what you think it will perform like in the future. Because what you think it will perform like with updates, might be completely wrong. Just like how it was completely wrong in the Bulldozer days.

And no... A single random guy showing a difference in performance between Windows 7 and Windows 10 does not confirm nor indicate any kind of performance improvement from updating a scheduler. Especially not since the Windows 7 scheduler should be equally unaware of Ryzen's special architecture as Windows 10 is.

Just because that one benchmark showed a 15% (or whatever it was) difference between 7 and 10 does not mean that it was caused by the scheduler (again, since the Windows 7 one should have the same issues as the 10 one has). It might be something completely unrelated, or an inaccurate benchmark.

Edited March 9, 2017 by wkdpaul
cleaned up

[Urishima]

57 minutes ago, Misanthrope said:

Bugs to neutralize eh?

 

Why does it works normally on Windows 7 which according to you shouldn't even work at all Mr. Bullshit Artist Microsoft Employee?

Actually, there is a possible (emphasis) explanation for it:

 

[cj09beira]

4 hours ago, LAwLz said:

Ehm... The conversation went like this:

 

Person: You get up to 20% performance improvement with SMT disabled so it won't be lower than that.

Me: Do you honestly think that we will get a 20% performance improvement?

Person: It will happen in some games!

Me: Got any examples of games where this will happen?

You: Games that uses less than 6 cores/8 threads will see this improvement.

Me: So you think most games will get a 20% performance increase?

You: I never claimed that! Use your brain! They will get between 10-20%!

 

If you follow the reply chain you will see that this is how the conversation went... 

 

 

Is that suppose to be me? Because I've been saying this since Ryzen reviews came out:

LGA 1151 - Still has no competition. If you were planing on getting one, then get one.

LGA 2011-3 - Dead. Ryzen killed it except for the very, very niche market (and things like servers, obviously).

Ryzen 7 - What you should get if you were thinking of getting something like the 6800K.

 

The only thing I have declared dead is X99, for most consumers at least.

 

I get that you want AMD to succeed and is scared of an Intel monopoly, but that's not excuse for buying into the hype machine or giving people hope that is just based on quite frankly ridiculous speculations.

 

Some first line support person from Microsoft tweeted that they will look into Ryzen? OMG Windows 10 scheduler confirmed for crippling Ryzen! Up to 20% gaming performance increase confirmed guys!

 

 

The reason why I bring up Bulldozer is because people said the exact same thing back then. They were always blaming something other than AMD and they kept getting disappointed over and over again. Lower your expectations and look at how Ryzen performs today, not what you think it will perform like in the future. Because what you think it will perform like with updates, might be completely wrong. Just like how it was completely wrong in the Bulldozer days.

And no... A single random guy showing a difference in performance between Windows 7 and Windows 10 does not confirm nor indicate any kind of performance improvement from updating a scheduler. Especially not since the Windows 7 scheduler should be equally unaware of Ryzen's special architecture as Windows 10 is.

Just because that one benchmark showed a 15% (or whatever it was) difference between 7 and 10 does not mean that it was caused by the scheduler (again, since the Windows 7 one should have the same issues as the 10 one has). It might be something completely unrelated, or an inaccurate benchmark

its a known fact that windows changed the scheduler after win 7 and for some reason the old version  reconizes ryzen just fine. True one guy isnt enough but its something that i would like to have tested.

Edited March 9, 2017 by wkdpaul
cleaned up

[Tataffe]

4 hours ago, LAwLz said:

The conversation went like this:

 

Person: You get up to 20% performance improvement with SMT disabled so it won't be lower than that.

Me: Do you honestly think that we will get a 20% performance improvement?

Person: It will happen in some games!

Me: Got any examples of games where this will happen?

You: Games that uses less than 6 cores/8 threads will see this improvement.

Me: So you think most games will get a 20% performance increase?

You: I never claimed that! Use your brain! They will get between 10-20%!

 

If you follow the reply chain you will see that this is how the conversation went... And you think I am the stupid one?

imo this proves how you constantly keep missing the point. @Morgan MLGman even highlighted stuff for you (which you conveniently left out of this summary, of course).

 

The original claim was an up to 20 % increase in performace in some games (even with references to why that would be), which you somehow interpret as "most games get 20 % performance increase". Additionally, you predict only small gains for the epic reason of "because [more gains] most likely will not happen" and not one piece of noteworthy proof.

Edited March 9, 2017 by wkdpaul
cleaned up

[3DOSH]

4 hours ago, LAwLz said:

Ehm... The conversation went like this:

 

Person: You get up to 20% performance improvement with SMT disabled so it won't be lower than that.

Me: Do you honestly think that we will get a 20% performance improvement?

Person: It will happen in some games!

Me: Got any examples of games where this will happen?

You: Games that uses less than 6 cores/8 threads will see this improvement.

Me: So you think most games will get a 20% performance increase?

You: I never claimed that! Use your brain! They will get between 10-20%!

 

If you follow the reply chain you will see that this is how the conversation went... 

 

 

Is that suppose to be me? Because I've been saying this since Ryzen reviews came out:

LGA 1151 - Still has no competition. If you were planing on getting one, then get one.

LGA 2011-3 - Dead. Ryzen killed it except for the very, very niche market (and things like servers, obviously).

Ryzen 7 - What you should get if you were thinking of getting something like the 6800K.

 

The only thing I have declared dead is X99, for most consumers at least.

 

I get that you want AMD to succeed and is scared of an Intel monopoly, but that's not excuse for buying into the hype machine or giving people hope that is just based on quite frankly ridiculous speculations.

 

Some first line support person from Microsoft tweeted that they will look into Ryzen? OMG Windows 10 scheduler confirmed for crippling Ryzen! Up to 20% gaming performance increase confirmed guys!

 

 

The reason why I bring up Bulldozer is because people said the exact same thing back then. They were always blaming something other than AMD and they kept getting disappointed over and over again. Lower your expectations and look at how Ryzen performs today, not what you think it will perform like in the future. Because what you think it will perform like with updates, might be completely wrong. Just like how it was completely wrong in the Bulldozer days.

And no... A single random guy showing a difference in performance between Windows 7 and Windows 10 does not confirm nor indicate any kind of performance improvement from updating a scheduler. Especially not since the Windows 7 scheduler should be equally unaware of Ryzen's special architecture as Windows 10 is.

Just because that one benchmark showed a 15% (or whatever it was) difference between 7 and 10 does not mean that it was caused by the scheduler (again, since the Windows 7 one should have the same issues as the 10 one has). It might be something completely unrelated, or an inaccurate benchmark.

I don't you think you fully get the picture. Ryzen is nothing like bulldozer so you cant really compare the two.
Ryzen CCX communication is tied to the Mem frequency which is limitation of Ryzen.
When the win 10 scheduler keeps switching/hopping threads this forces more CCX communication thus impacting performance a lot since the bandwidth is not that great. 
Win 10 scheduler is not the source of the problem but its one of the causes.
The proof for this is Ryzen scales a lot  better with higher Mem frequency's than Intels archs and we see improvements in benchmarks that don't really depend on Mem frequency's such as Cinebench.

Edited March 9, 2017 by wkdpaul
cleaned up

[cj09beira]

5 minutes ago, 3DOSH said:

I don't you think you fully get the picture. Ryzen is nothing like bulldozer so you cant really compare the two.
Ryzen CCX communication is tied to the Mem frequency which is limitation of Ryzen.
When the win 10 scheduler keeps switching/hopping threads this forces more CCX communication thus impacting performance a lot since the bandwidth is not that great. 
Win 10 scheduler is not the source of the problem but the one of the causes.
The proof for this is Ryzen scales a lot  better with higher Mem frequency's than Intels archs and we see improvements in benchmarks that don't really depend on Mem frequency's such as Cinebench.

And in games it gets worse because in games threads depend a lot in one another and if thoses threads arent in the same ccx it causes problems as we are seeing

[CTR640]

Microbugs 10 at it again, nerfing Zen so it makes Zen look shit compared to Crinchtel.

[Mira Yurizaki]

I'm curious what's different about the scheduler in Windows 10 vs. Windows 7 that supposedly makes Ryzen worse. The only thing I'm aware of is that in Windows 8, Microsoft increased the default event polling period from something like a handful of miliseconds to somewhere in the neighborhood of 50-70ms and implemented deferred events so that everything gets handled on the next period.

 

But I doubt this would have an effect in games since applications can request lower the polling period (like what Chrome did).

 

EDIT: Oh, nevermind. This is probably a case of Microsoft not updating their scheduler to have a case for SMT when the CPUID is an AMD one.

 

Code paths, gotta love them, gotta hate them.

[LAwLz]

3 hours ago, cj09beira said:

its a known fact that windows changed the scheduler after win 7 and for some reason the old version  reconizes ryzen just fine. True one guy isnt enough but its something that i would like to have tested.

Oh, it's a known fact you say? Then please, explain this known fact to me. Because if Ryzen requires such careful tuning it really doesn't make sense to me that a 7 year old OS will understand it just fine and be optimized for it, yet an OS that's about 1.5 years will not.

I mean, let's ignore that Windows 10 is notorious for just being inconsistent in general. Let's blame it on the scheduler even though 99.9% of people who repeat this probably doesn't understand the first thing about how a scheduler works.

So tell me, what evidence do you base this belief on?

 

3 hours ago, Tataffe said:

imo this proves how you constantly keep missing the point. @Morgan MLGman even highlighted stuff for you (which you conveniently left out of this summary, of course).

 

The original claim was an up to 20 % increase in performace in some games (even with references to why that would be), which you somehow interpret as "most games get 20 % performance increase". Additionally, you predict only small gains for the epic reason of "because [more gains] most likely will not happen" and not one piece of noteworthy proof. That is on the border of trolling.

I am not missing the point... 

What you are referring to as "missing the point" was me abridging his posts. The reason why I didn't include that specific wording is because his original post was quite frankly *badly* worded. You can read it in two ways.

You can read it as "we will see a 20% performance increase, minimum" which is how I chose to interpret it. Or you could interpret it word for word, in which case his post is meaningless. Why is it meaningless? Because his post says "the performance increase will be between 0 and infinity". His posts says "up to 20%, or more". That means it could be 1% increase, and it would still be included because of the "up to" portion, and it could be 1000% and it would be included in the "at least" portion.

I chose to interpret it the way I did because:

1) I assumed he didn't intend to say "the performance increase could literally be anything, it might not even increase performance at all".

2) Right after he had mentioned the 20% increase he said "it won't be lower than this value". If someone says "I will give you up to 20 dollars, but not less" then I think it's rational to assume that you meant "not less than 20 dollars".

 

 

I thought this was stupid, hence why I asked for more specific examples of games which should see this alleged 20% increase (which seems taken out of thin air). When I asked for examples, another person jumped in and said "any game that uses less than 6/8 threads".

 

That's how I "interpreted his post as most games gets 20% performance increase". Because if you follow the conversation that was literally what was said. Hence why I even asked several times is that was what he actually meant. Because I knew that he was misreading things.

Edited March 9, 2017 by wkdpaul
cleaned up

[LAwLz]

46 minutes ago, M.Yurizaki said:

EDIT: Oh, nevermind. This is probably a case of Microsoft not updating their scheduler to have a case for SMT when the CPUID is an AMD one.

Wouldn't Windows 7 also have the same issues then?

 

I don't know how Windows' scheduler works under the hood, but in Linux (from my limited understanding) all SMT is handled more or less the same. You just have to assign a unique core ID to each core, and then set the ThreadID to be unique for each thread. The scheduler will see that two threads has the same core ID, and it will figure out that it's an SMT enabled core and assign tasks accordingly.

I assume that Windows does it the same way. If that is the case, then you don't need any specialized code to support Ryzen, or any SMT implementation really. If it works on let's say Intel, then it will work on AMD too.

 

The Linux patch was needed because each thread got their own core ID as well. It was like ~5 lines of code to fix. If this is what is happening in Windows (which I assume Microsoft will check the first thing they do because of the rumors floating around) then someone could make a patch which fixes it in like 10 minutes.

[Mira Yurizaki]

16 minutes ago, LAwLz said:

Wouldn't Windows 7 also have the same issues then?

Yeah, I thought about that a few minutes later, but couldn't really think of anything to say so I just left it there. But now I have!

 

It is possible that Windows 7 and prior tries to schedule everything on the lower numbered logical processors until they can't accept anymore work, whereas Windows 8 or 10 tries to load balance everything on all logical processors. If that's the case, then Windows 7 is shoving everything onto on CCX, avoiding cross CCX communication by a coincidental fluke. I can see merits and drawbacks to either, but that's the only explanation I can think of.

Quote

I don't know how Windows' scheduler works under the hood, but in Linux (from my limited understanding) all SMT is handled more or less the same. You just have to assign a unique core ID to each core, and then set the ThreadID to be unique for each thread. The scheduler will see that two threads has the same core ID, and it will figure out that it's an SMT enabled core and assign tasks accordingly.

I assume that Windows does it the same way. If that is the case, then you don't need any specialized code to support Ryzen, or any SMT implementation really. If it works on let's say Intel, then it will work on AMD too.

 

The Linux patch was needed because each thread got their own core ID as well. It was like ~5 lines of code to fix. If this is what is happening in Windows (which I assume Microsoft will check the first thing they do because of the rumors floating around) then someone could make a patch which fixes it in like 10 minutes.

It would make Linux to try to do this as generically as possible, since it's supposed to work on a multitude of architectures. However I wouldn't put it past Microsoft to only focus on x86 and as far as Windows is concerned, a logical processor is a logical processor.

 

But is the issue really just that Windows 10 has a problem with AMD's SMT, or that Windows 10 is unaware of relatively low bandwidth between the CCX's L3 caches?

 

EDIT:

I found an old article but it's probably relevant considering kernel changes to established components are rare:

Quote

In a symmetric multiprocessing (SMP) system, the dispatcher selects a processor to run a thread. The dispatcher first selects the thread's ideal processor. If the ideal processor is unavailable, the dispatcher selects the processor the thread ran on previously. If neither processor is available, the dispatcher selects another idle processor or a processor running a lower-priority thread.

 

And there's another bit on the MSDN website:

Quote

When you specify a thread ideal processor, the scheduler runs the thread on the specified processor when possible. Use the SetThreadIdealProcessor function to specify a preferred processor for a thread. This does not guarantee that the ideal processor will be chosen but provides a useful hint to the scheduler. On systems with more than 64 processors, you can use the SetThreadIdealProcessorEx function to specify a preferred processor in a specific processor group.

 

[LAwLz]

Just now, M.Yurizaki said:

It is possible that Windows 7 and prior tries to schedule everything on the lower numbered logical processors until they can't accept anymore work, whereas Windows 8 or 10 tries to load balance everything on all logical processors. If that's the case, then Windows 7 is shoving everything onto on CCX, avoiding cross CCX communication by a coincidental fluke. I can see merits and drawbacks to either, but that's the only explanation I can think of.

I doubt that's how Windows 8 and 10 does it. It doesn't make sense for so mobile focused OSes to try and spread load out over all cores as much as possible, because that is really bad power efficiency wise. You want as few cores to be fired up because of core parking and turbo boost, and I can't really think of any benefit of doing it.

 

To me it seems like the people trying to rationalize this whole "Windows 10's scheduler is to blame!" are working backwards. They came up with a conclusion (that it's the scheduler that's bad) and is now trying to find evidence to support their predefined conclusion.

You collect evidence first and then draw a conclusion. You don't look for evidence and reasoning to support your conclusion.

 

4 minutes ago, M.Yurizaki said:

It would make Linux to try to do this as generically as possible, since it's supposed to work on a multitude of architectures. However I wouldn't put it past Microsoft to only focus on x86 and as far as Windows is concerned, a logical processor is a logical processor.

Well I wouldn't be surprised if Microsoft designed Windows in a really terrible way, but I can't really think of any reason why they would do it. Why assume that they have designed it in a really convoluted way when there is a really simple and efficient solution for it?

 

10 minutes ago, M.Yurizaki said:

But is the issue really just that Windows 10 has a problem with AMD's SMT, or that Windows 10 is unaware of relatively low bandwidth between the CCX's L3 caches?

Who knows? I am not even sure that it is a scheduler issue. I haven't seen any evidence to support that. Seems like this all stems from a single guy posting a screenshot showing Windows 7 beating Windows 10 in a handful of benchmarks (might even just have been one benchmark) and then people jumped to the conclusion that it must be Windows 10's scheduler (because surely nothing else has changed between 7 and 10, right?).

[Morgan MLGman]

4 hours ago, LAwLz said:

Such as? I highly doubt we will see a 20% performance increase (or more) in more than one or two games, and chances are those gains won't be from a scheduler update since it's not like a bug like this supposedly is will only affect certain games to that degree.

See, there are a few games where the performance impact lost by having SMT enabled is quite significant, I said 20% from the top of my head but judging by the benchmarks of Total War: Warhammer (IIRC) the difference was just around that mark, of course those games aren't EVERY game but I've seen a review yesterday which tested SMT ON/OFF setting in 15 games and having it off actually improved average performance by 6%, treating the 1800X as a regular 8 core CPU. Considering that enabling SMT should help quite a bit in most newer games, the average performance uplift across the board should be quite significant in my opinion...

4 hours ago, LAwLz said:

This is the case with Hyperthreading too. I haven't checked it out in recent years, but even with Ivy Bridge you could find programs that performed worse with Hyperthreading on compared to off. That was not a Windows scheduler issue, since the scheduler was already Hyperthreading aware.

You misread I quoted myself saying that turning SMT off improved performance in some games to prove a point and said below the quote that there also are games where turning it off lowered performance, just as it usually is, a good example of that is the difference in framerates and most importantly frametimes between newest i5s and i7s.

 

I'm following the CPU market quite closely since the Haswell launch and haven't heard of any noticeable performance improvements with HT off on Intel's i7 chips. This technology is here for so long that it should not lower the performance in any way to be fair

4 hours ago, LAwLz said:

High-performance power plan improves performance on all platforms... It's not a Ryzen specific "issue". That's how it is meant to work.

Yeah, the problem is that the setting is not optimized for Ryzen and it just doesn't work as intended as even with heavy workloads, you could have issues maintaning base clocks if High Performance power plan was disabled due to AMD's SenseMI technology interfering with it. Though that's fairly simple to fix, it's an issue but if configured properly, the system can still work as intended but most of the initial reviewers probably didn't know about it thus very varied results.

 

I highly suggest you take a deeper look into the scheduler issue if you're interested, there definitely is one, though to what extent and how "fixable" it is, is unknown to me.

However logic suggests that the performance uplift in those specific games where SMT OFF improved performance should be at least the equal to the % lost with SMT ON after the fix rolls out.

 

So far I've read that the W10 scheduler can't differentiate physical cores from logical ones on the R7 lineup yet, thus lower performance in some instances where it puts the singlethreaded workload on a weaker logical core instead of a faster physical one, and that it "treats" the Ryzen R7 as a dual 4C/8T CPU due to the fact that it has two CCX's (Core Complexes) with four cores each, and if the workload is spread among let's say two cores that lie on CCX0 and CCX1 each, the data transfer doesn't go through the L3 Cache which has bandwidth of just under 200GB/s but through a high-speed interconnect called "AMD Infinity Fabric" which has around 22GB/s of bandwidth.

That creates a HUGE bottleneck and results in increased latency which was proven through extensive testing of some tech sites such as techpowerup.com, meaning that if it's even possible to fix that (cause I genuinely don't know, I'm not an AMD engineer ), the performance uplift should be quite significant.

 

It's not like I have huge expectations, but it's definitely interesting that there are games where the 1800X performs really well and sometimes outperforms the 6900K and even there are rare instances where it somehow beats the 7700K by a tiny bit, but there are games where it completely fails to deliver and that heavily indicates on an optimization issue cause it performs way below its specs and expected performance. This is what I'm hoping for, that it's gonna get fixed eventually and Intel starts getting some real competition and a run for their money. With lower core count and most likely higher clock speed chips rolling out fairly soon the gaming part of the market should also be satisfied if the potential issues I mentioned above get fixed.

 

Sorry for a long post, this stuff is just exciting to me.

[Mira Yurizaki]

14 minutes ago, LAwLz said:

I doubt that's how Windows 8 and 10 does it. It doesn't make sense for so mobile focused OSes to try and spread load out over all cores as much as possible, because that is really bad power efficiency wise. You want as few cores to be fired up because of core parking and turbo boost, and I can't really think of any benefit of doing it.

Frequency is a huge power hog however, and that depending on the circumstances, it may be better to keep more cores up, but in a light sleep state that can wake up immediately, than it is to have fewer cores fired up, only to realize you need to fire up one more. And since it probably takes less voltage to run a processor at a lower speed, you can save power there as well.

 

14 minutes ago, LAwLz said:

Well I wouldn't be surprised if Microsoft designed Windows in a really terrible way, but I can't really think of any reason why they would do it. Why assume that they have designed it in a really convoluted way when there is a really simple and efficient solution for it?

I'm not saying they designed it in a convoluted way, that only they focused on x86. And if anything, they probably recycled their SMP scheduling logic.

[Tataffe]

1 hour ago, LAwLz said:

I am not missing the point... 

What you are referring to as "missing the point" was me abridging his posts. The reason why I didn't include that specific wording is because his original post was quite frankly idiotically worded. You can read it in two ways.

You can read it as "we will see a 20% performance increase, minimum" which is how I chose to interpret it 

Here it is for you to read again:

9 hours ago, Morgan MLGman said:

Well, considering that with SMT disabled you get up to 20% performance more in some games then it won't be lower than this value. So it's pretty much: any performance gains from having SMT disabled in specific games will be the lowest value that the scheduler fix should bring for the SMT-on setting.

I'll try to put it in simple words: "Some games benefit from having SMT disabled. The gains we have observed range from only little to about 20 %. I assume that fixing the scheduler will bring at least the same gains for each game as disabling SMT."

 

I see no other way to read this than the one you listed second. Especially since the first way (which you picked due to your outstanding reading comprehension skills, I assume) would actually make him look like an idiot. And no, his post was not "idiotically worded". You misread and then failed to recognize that your interpretation doesn't make sense imo.

 

Also you still have to prove your statement about only little gains, may I remind you.

Edited March 9, 2017 by wkdpaul
cleaned up

[Guest Kloaked]

2 hours ago, CTR640 said:

Microbugs 10 at it again, nerfing Zen so it makes Zen look shit compared to Crinchtel.

Congratulations! You've won! Please click here to claim your prize!

[SpaceGhostC2C]

31 minutes ago, Morgan MLGman said:

 

I'm following the CPU market quite closely since the Haswell launch and haven't heard of any noticeable performance improvements with HT off on Intel's i7 chips. This technology is here for so long that it should not lower the performance in any way to be fair

Not to detract from your argument in any way, since you are discussing gaming here, but a small note: in parallel computing applications (a task far away from gaming) disabling HT is known to improve performance, and the rather surprising result is that it's no longer so clear-cut. This is because when you are saturating your cores with identical work (and need to gather the results from all the cores later, which adds overhead) it is counterproductive to run more threads than cores anyway, and running as many threads as cores with HT enabled leads the OS to believe that there are "idle logical cores" to assign to any garbage running in the background (I tested this mostly in Debian, but I can only assume it will be worse in Windows ). Disabling HT makes sure the CPU is seen as 100% busy by the OS, and only high priority stuff can interrupt.

A recent test with my laptop did not suffer from HT, even managed to get better results by doing more threads than cores, so maybe I'm not as good parallelizing anymore   

OK, sorry, no more old man digression  

[WkdPaul]

Thread cleaned (I think ?)

 

Please keep in mind the Community Standards;

 

• Ensure a friendly atmosphere to our visitors and forum members.
• Encourage the freedom of expression and exchange of information in a mature and responsible manner.
• "Don't be a dick" - Wil Wheaton.
• "Be excellent to each other" - Bill and Ted.
• Remember your audience; both present and future.

[LAwLz]

2 hours ago, Morgan MLGman said:

of course those games aren't EVERY game but I've seen a review yesterday which tested SMT ON/OFF setting in 15 games and having it off actually improved average performance by 6%, treating the 1800X as a regular 8 core CPU.

Source?

~6% is a lot more in line with what I would expect from this, and it's a long way from the 10-20% claimed in this thread.

 

2 hours ago, Morgan MLGman said:

You misread I quoted myself saying that turning SMT off improved performance in some games to prove a point and said below the quote that there also are games where turning it off lowered performance, just as it usually is, a good example of that is the difference in framerates and most importantly frametimes between newest i5s and i7s.

Sorry, quoted the wrong part of your post. What I meant to say was that there are programs which performs worse with hyper threading enabled too. At least they used to back when I was looking into it with Ivy Bridge.

 

2 hours ago, Morgan MLGman said:

I'm following the CPU market quite closely since the Haswell launch and haven't heard of any noticeable performance improvements with HT off on Intel's i7 chips. This technology is here for so long that it should not lower the performance in any way to be fair

I can't find any article looking into it later than Ivy Bridge, but here is the article.

3 out of 8 benchmarks performed better with hyperthreading off. The benefit was much smaller than the drawbacks though so it was still much better to leave it on.

I don't know how Intel has tweaked hyperthreading over the years, but the article makes it sound like they are the ones responsible for making sure it performed like it should, not the OS vendors.

Quote

The implementation of this technology found in Ivy Bridge processors, and to a large extent in Sandy Bridge processors, is quite robust and well adapted to our current crop of OS schedulers.

[...]

As performance enhancing technologies go, Hyper-Threading is surely one of the more mature examples that Intel has to offer.

 

 

2 hours ago, Morgan MLGman said:

Yeah, the problem is that the setting is not optimized for Ryzen and it just doesn't work as intended as even with heavy workloads, you could have issues maintaning base clocks if High Performance power plan was disabled due to AMD's SenseMI technology interfering with it. Though that's fairly simple to fix, it's an issue but if configured properly, the system can still work as intended but most of the initial reviewers probably didn't know about it thus very varied results.

Any reviewer worth a damn will use high performance mode, and I don't buy the argument that a lot of reviewers didn't know because AMD was telling the reviewers to uses high performance mode.

 

2 hours ago, Morgan MLGman said:

I highly suggest you take a deeper look into the scheduler issue if you're interested, there definitely is one, though to what extent and how "fixable" it is, is unknown to me.

How do you know? I am looking into it, and I can't find any evidence that that's what is actually going on. I see a lot of people guessing that it's the issue, but nobody with any evidence to support that theory.

 

2 hours ago, Morgan MLGman said:

However logic suggests that the performance uplift in those specific games where SMT OFF improved performance should be at least the equal to the % lost with SMT ON after the fix rolls out.

Why? That wasn't the case with Intel back in the Ivy Bridge era, and I don't see why it should be true here.

 

 

1 hour ago, Tataffe said:

Also you still have to prove your statement about only little gains, may I remind you.

That's not how arguments work...

There were really high claims in this and other thread being thrown around left and right, and I said people should lower their expectations.

I can not prove something that might happen in the future, and neither can the opposite side. That is why I have simply been recommending people to keep their expectations low. Did I ever claim to know which kind of performance increases we will get? I am only aware of two people who has made statements about what kind of performance we should expect, and it's the two people I have told to lower their expectations because we don't even know if there is a scheduler issue at all to begin with.

[Doobeedoo]

Well, hopefully they'll patch it soon enough so we can get this issue over with.

 

[cj09beira]

2 hours ago, LAwLz said:

 

this video is a good watch,

 

[AnonymousGuy]

Hey AMD, maybe it would be a good idea to do basic OS validation BEFORE launching.  Just a suggestion.

[leadeater]

3 hours ago, AnonymousGuy said:

Hey AMD, maybe it would be a good idea to do basic OS validation BEFORE launching.  Just a suggestion.

Rather hard if the OS developer goes "Meh I don't care, we don't want to spend any time helping you". Not that Microsoft did that or anything but AMD can only control so much. And no there is not a hardware solution to this problem, Microsoft spent many years getting Intel Hyper-Threading to where it is now and they will need to do it again for AMD with the bonus of being able to do it much quicker from previous experience (if they bother).