[News] Microsoft Windows Bug Is Holding Back AMD Ryzen

[StyM]

http://www.guru3d.com/news-story/mic...amd-ryzen.html

We've been addressing the topic widely in our reviews, it makes little sense for Ryzen processors that are that powerful to not perform as well in games in CPU bound situations. Microsoft now kind of confirms the issue that Windows 10 does not detect the simulated SMT threads properly, and the simulated SMT threads are weaker compared to real CPU cores. This could hinder the game performance.

In nothing more then a tweet Microsoft seems to have confirmed the existence of this issue. So basically each core gets an extra simulated thread, that is SMT (simultaneous multi-threading) or what Intel refers to as hyper-threading. A real core obviously is stronger compared to a virtual threaded one. Aka we are back to thread-scheduler bugs. Things remain a bit trivial we feel, when we disabled SMT in the BIOS (and just ran 8-threads) performance only increased marginally on some titles, with an exception for some. But who knows, titles like Rise of the Tomb Raider (which is hit the most) might benefit from properly threaded processors big-time.

There's more going on with the scheduler though, a weird fact exposed itself that the output of the diagnostic tools for the Windows 10 scheduler that distributes these computing tasks across cores, estimated an overly optimistic cache memory per core for Ryzen. A Ryzen 7 CPU seems to be recognized as a processor with a 136 MB combined cache, where it is in reality is obviously is 20 MB for L2 and L3 caches.

Meanwhile Windows 10 RTC seems to be disliking 0.25x multipliers, causing timing issues and Microsoft?s scheduler might not understand the different CCX core-to-core latencies either. So yes, most paths now lead back to the thread scheduler.

In short, the Windows thread-scheduler does not line up with the AMD Ryzen architecture.

Back in the days when Intel introduced Hyper-threading many similar issues have been addressed. From here on-wards it will be a waiting game as we'll have to wait and see how Microsoft will address this matter and see what effect these two factors will have on the slightly disappointing game performance results. But sure, it is terrific news that Microsoft found an issue and is correcting it, let us hope that Microsoft sorts out the scheduler and cache issues as soon as possible. Maybe the upcoming Patch Tuesday update scheduled for March 14 will bring this coveted patch.

[AliG]

I dunno that it's that big of a difference though, these tests show an average of just +3% after disabling SMT:
http://wccftech.com/ryzen-gaming-benchmarks-roundup/

The cache issue may be a larger effect, but I would assume you'd see problems across the board if you were truly freuqeutnly overflowing the cache

[Ket]

In my own testing for the Gigabyte Gaming 5 review I'm doing memory latency with any application that has any kind of reliance on it stomps down performance so badly people in the next universe are able to see it. Heres a few bits from the review article of mine;



Memory performing as it should be @ 3200MHz, 15-15-15-35, 80ns latency



When the memory latency bug hits (300ns+).

[Sparky]

So the question is how does the latency get fixed? Zen2?

[NEOAethyr]

So the question is how does the latency get fixed? Zen2?

No not zen2 I think, because that would entail a whole redesign, they worked from the ground up on this interconnect module thing.
So we're stuck with this issue for at least 5 years.
The only way to get around it is to limit games to 8 threads..., which is the opposite of what amd is actually saying, they act as if games using 16 threads would help... (they're essentially lying...).
As for treating it as a numa node thing, could help, but as it stands, in windows, if you do this, then no app can use all 16 threads, meaning thee goes the media encoding perf and other workstation needs...

However, if in zen2, they decided that instead of 4 cores per module/ccx, they went 8 cores per, and ensured the ccx was 128bit wide..., then all the issues are fixed, no soft updates needed.
Except some minor smt issues that intel has a prob with as well.

I don't know, whatever they do I hope they doit quick.
So I can cautiously recommend this thing already :\.
I've been trying to convince my roomate this will be a good upgrade for him, but after this fiasco, he's dead set on intel.
Another dude that has been waiting for this for years to upgrade an old amd setup with nv 1080, I had to call him up and tell him it's a peace of junk.
Which I don't really like doing...
Right now as is, if they can do at least a little bit of work on the scheduler, then I could recommend it as a half @ssed workstation, and gamer platform (it games good enough for now, by the time it's a major issue zen2 will be out)

[Sparky]

Fiasco, piece of junk... A little harsh, perhaps?

[Charles Wirth]

I think it can be fixed in BIOS update

[MattiasNYC]

Maybe I'm misunderstanding this, but;

Isn't the 'casualty' here happening when data isn't found in any of the caches in the CCX in which the thread lives? If that's the case, then the "solution" could be as "simple" as making sure the threads don't migrate from CCX1 to 2 or vice versa. The latency would still be "the same", but it wouldn't be an issue per se because as long as the relevant data is found in a 'local' cache we're avoiding it.

Or am I not getting this correctly?

[chew*]

I think 3 months from now as I feel that is a realistic amount of time to get everything in all aspects sorted people are going to see this chip in a whole new light.

[AliG]

I think 3 months from now as I feel that is a realistic amount of time to get everything in all aspects sorted people are going to see this chip in a whole new light.

I think it'll still need a respin of the silicon to get a bit more overclocking headroom, but I agree that eliminating the MS bugs and reducing the memory latency via bios updates should help a lot.

[Ket]

I think it'll still need a respin of the silicon to get a bit more overclocking headroom, but I agree that eliminating the MS bugs and reducing the memory latency via bios updates should help a lot.

Craptacular UEFI that isn't even out of the alpha stages aside on the Aorus Gaming 5 I'm reviewing (I've never had a board thats more of a chore to review than this thing) my plain 1700 has shown itself to have some decent headroom. Had it up to 4GHz so far.

[drmrlordx]

I dunno that it's that big of a difference though, these tests show an average of just +3% after disabling SMT:

The cache issue may be a larger effect, but I would assume you'd see problems across the board if you were truly freuqeutnly overflowing the cache

Disabling SMT doesn't stop the scheduler from bouncing threads from one CCX to the other. That causes all kinds of problems: cache thrashing, etc.

Stop moving threads from one CCX to the other and problem solved. The SMT issue is secondary, though it should be solved via scheduler update as well. MS just needs to get on the ball.

[chew*]

Exactly the cache is getting hammered.

Time will tell.

While it won't overclock better ( time will tell on that as process matures ) it can perform better and these are just bugs and software optimizations that need to be done.

There is still performance to be gained elsewhere on the hardware side of things.

Thats been proven already at the cost of high pci which i have no clue why its linked but it wont matter once we gain access to hidden timings as much.

We are just proving the point that theres alot left on the table when you trick the timing tables.

I would definitely not call it junk though. It made a decent showing with all this going against it...imagine when its fixxed.

[vario]

First of all, how did they not see this problem coming is beyond me.Im pretty daft about all of it, and i had a "wtf, thats gonna be problem" moment the day i saw that there are in fact two L3 caches on one chip.
Second of all.
This thing is blown outta proportion.
So 1800X is on average like ~15% slower than the newest intels in gaming workloads.
a) You wont see any difference unless you game with 1080 and up.
b) Mostly even then only in 144hhz/sli setups. which are like 10% of all at most.
c) Its still gonna be like 15%! its nothing earth shattering
d) 1700 costs ONE THIRD of 6900K and does 90% of the work (OCed) .
e) there are already quick fixes of sorts to get back up to half of those 15% on average!

To me this whole gaming thing is mostly a non issue, yes for specific people it is, for most, not.
BUT, mainboards and bioses. NOT READY.

[chew*]

Mainboard bios's are at the mercy of the agesa code.

How many vendors have 5 adjustable timings unlocked? Um all of them. You think thats a coincidence?

Anything across all vendors at this point = agesa

While some seem a bit barren and lackluster you need to look at it from a broader aspect. The agesa is not finalized still being worked on. A simple functional bios takes less work more reliable less manpower. When agesa is more polished and will be around for awhile then you throw the bells and whistles at it.

And ill fix this for you. You wont notice a difference in gaming unless you use 1080p and lower resolutions.

[Ket]

Mainboard bios's are at the mercy of the agesa code.

How many vendors have 5 adjustable timings unlocked? Um all of them. You think thats a coincidence?

Anything across all vendors at this point = agesa

While some seem a bit barren and lackluster you need to look at it from a broader aspect. The agesa is not finalized still being worked on. A simple functional bios takes less work more reliable less manpower. When agesa is more polished and will be around for awhile then you throw the bells and whistles at it.

And ill fix this for you. You wont notice a difference in gaming unless you use 1080p and lower resolutions.

Myself and Fugger would beg to differ Neither of us are particularly impressed with the Gaming 5, Fugger less so than even me at this stage. All the other manufacturers I have seen have a far more feature complete UEFI regardless of being at the mercy of the Agesa and some of the issues with the AX370 Gaming 5s UEFI are just downright amateurish.

[chew*]

Right now as it stands....its simpler to overclock than sandybridge was. There is multi and voltage.

K.i.s.s applies quite well here. Boring yes i will say that.

[Ket]

Right now as it stands....its simpler to overclock than sandybridge was. There is multi and voltage.

K.i.s.s applies quite well here. Boring yes i will say that.

Simple is one thing, the Aorus UEFI is downright barren! There isn't even options such as to disable on-board audio, and you must of noticed the same options being displayed in different menus along with apparently redundant naming for some settings in the UEFI such as CPU LLC: Normal, Standard, Medium. Three settings, all essentially with the same meaning. Things like that are such obvious mistakes they should not of made it past the first few alphas never mind in to a UEFI thought good enough to go with a retail product.

[chew*]

Llc did you test each one? Did it change? I did and it did. So functional first and foremost.

The redundant crap is really not an OMG they put an option there twice. kill them..

Me personally i have focussed solely on function. Thats what i report. Timing not applying = problem

Two locations to tick manual timing not a concern as long as not causing conflicts.

Disabling devices? They might not be able to. I have not confirmed it yet but requested it.

[sergiu]

Watching the thread with interest. Given how obvious are the implications of CCX groups, it's hard to imagine Microsoft was not aware of those and does not have a scheduler version that works. One could argue Microsoft might have been paid to delay such a fix until after launch.
One big question that nobody asks: how does Linux kernel handle Ryzen?

[Ket]

Llc did you test each one? Did it change? I did and it did. So functional first and foremost.

The redundant crap is really not an OMG they put an option there twice. kill them..

Me personally i have focussed solely on function. Thats what i report. Timing not applying = problem

Two locations to tick manual timing not a concern as long as not causing conflicts.

Disabling devices? They might not be able to. I have not confirmed it yet but requested it.

I don't mean to sound like a arse here but a bug is a bug, pure and simple. It does not matter how minor it is and it should be being reported so it can be addressed. The same applies to duplicate entries, it is amateurish and looks thoroughly unprofessional never mind looking lazy like somebody really couldn't give a rats arse. Know why the Asrock P67, Z68 and Z77 UEFIs were / are so good? Because I told my contacts about everything, not mentioning things is what leads to a poor product. Redundant options also unnecessarily confuse matters, not to mention once again look amateurish, sloppy, and lazy.

Bottom line: Every bug, glitch, minor issue, impedes the functionality and / or usability in some way.

[chew*]

Tbh i dont care how good something looks. I care how it works. Function over form always.

If normal medium and standard llc all did the same amount of compensation....problem.

I dont care if they call it noob,amateur,novice,pro,1337 as long as each lvl works.

If i reported that stuff as a bug i would end up on an ignore list.

Bugs cause things to either not function or break. That is what bug reporting is.


If it looks cool does not work its a brick...

Go buy a few more boards in the same budget range.

Not top tier budget as 5 series has always been nothing more.

Lmk what you think.

Im realistic about my expectations and i read specs before i buy.

I have been reading them all night looking for a spare board. For instance msi tomahawk and carbon both do not list my exact memory as qvl. Both support "3200 oc"

If i buy it and it does not work guess who is at fault?

I also can see that bclock is not adjustable on a variety of boards by doing this.

So when i complain about a board i make sure its worth complaining about and i definitely do not complain about the lack of a feature that was never advertised in the first place.

Newsflash the asrock boards that are badass?...nick shih. Ill leave it at that

[NEOAethyr]

I'll complain just about anything when it comes to bios'es.
Though I try to prioritize when I might ask for because it's essentially begging and you're not likely gonna get it anyways...
You get what you pay for when it comes to mobo's, if one has all the ios options and the other don't, pretty obvious you should get the one with the better bios...
And board design of course...

I personally wouldn't go gigabyte, in the past I've heard nothing but bad things about there bios.
Same goes for msi and evga.
However that could change, they don't make some decent looking boards sometimes.
And from what I've seen on the msi's (very little myself), they seem to be adding a bunch of little efi programs which is cool (Don't like the networking idea though).

I've heard good things about the asrocks, but I've never touched one personally.
That would be there highest end one's...
And of course asus.

We don't have dfi, abit or epox anymore.
Dfi only makes oem crap boards now.

And there's the odd generic one, biostar, which sometimes looks like crap, and sometimes it's ok and can be modded to high heaven... (it can be made to be "ok" lol...)


As for ryzen and this bug thing.
I didn't mean to imply that it was a complete disaster....
They'll make it in the end, everything is cool .
However, I am disappointed.
They put some marketing crap over making a cpu for pc's.
All so they could swap modules to make products no one knows right now that anyone will buy....

Take there arm apu for ex., if it ever see's the light of day...
I used to think, heck why not, maybe samsung will buy it, or maybe samsung will buy some tegra chips.
Or whatever, amd on the pi for ex...

But now thinking about it, I don't know that much but from what I gather, amd's drivers aren't that good on linux in general...
People have to resort to open src because they are so bad.
And those are like half the speed of the windows drivers I guess.

Ah crap I gotta get already, anyways ryzen isn't a 100% fail, but the design sucks (my opinion)...

[Ket]

Tbh i dont care how good something looks. I care how it works. Function over form always.

If normal medium and standard llc all did the same amount of compensation....problem.

I dont care if they call it noob,amateur,novice,pro,1337 as long as each lvl works.

If i reported that stuff as a bug i would end up on an ignore list.

Bugs cause things to either not function or break. That is what bug reporting is.


If it looks cool does not work its a brick...

Go buy a few more boards in the same budget range.

Not top tier budget as 5 series has always been nothing more.

Lmk what you think.

Im realistic about my expectations and i read specs before i buy.

I have been reading them all night looking for a spare board. For instance msi tomahawk and carbon both do not list my exact memory as qvl. Both support "3200 oc"

If i buy it and it does not work guess who is at fault?

I also can see that bclock is not adjustable on a variety of boards by doing this.

So when i complain about a board i make sure its worth complaining about and i definitely do not complain about the lack of a feature that was never advertised in the first place.

Newsflash the asrock boards that are badass?...nick shih. Ill leave it at that

Firstly, I'm not having a go at you, I'm saying Gigabyte need to be informed of all the issues the Gaming 5 UEFI has as in this regard they are trailing their competition and by no small margin. I'm also not the only one whos had time with the Gaming 5 to say these things either, I've just said them in more detail to help Gigabyte isolate and fix said issues as efficiently and as quickly as possible. It would be appreciated if you would pass on the detailed list of bugs I have found, a list of them is here. If not, I will be attempting to reach out to Gigabyte anyway.

1. CPU LLC: How are you measuring that each and every level makes a difference? At the very least each option need renaming so it makes coherent sense. Eg; Standard, Normal, Medium > Level 1, Level 2, Level 3.
2. I am absolutely certain Gigabyte do not want to be taken for amateurs, so your argument is just unfathomable here.
3. Devs need to know about any and all bugs in order to address them in order of priority, that is their job. If you think Gigabyte are going to put you on a ignore list for reporting bugs.. well that says a lot about Gigabyte doesn't it?
4. Bugs are all shapes and sizes, not what just doesn't work.
5. What? I didn't even bring up board looks, everything I've mentioned is a functionality or feature issue.
6. I will be asking around contacts for more AM4 boards to look at, and I will specifically ask if theres any I can have in the same price range as the AX370 Gaming 5, which for the record in the UK is retailing for anywhere between 170-230. The AX370 Gaming 5s current poor UEFI doesn't make it worth that.
7. Theres nothing unrealistic about expecting a UEFI to be roughly on-par with competitor alternatives.
8. Everyone knows all AM4 boards are having a lot of issues, but thats no reason to not at least make the UEFI as good as possible while AMD work on things their side.
9. Nobody has said the AX370 Gaming 5 doesn't work, just that the UEFI needs a lot of work to bring it up to scratch.
10. Not sure what the bclock comment has to do with anything as it wasn't bought up, AX370 Gaming 5 doesn't have a clock generator thats pretty much the end of that. Not a deal breaker.
11. I'm pretty sure when a UEFI lacks pretty basic features and has alpha stage mistakes still in something deemed ready for retail anyone has a right to call it out.
12. I did not mention ALL Asrock boards, nor did I say any of them are what LN2 folks would want to use, I said the P67, Z68, and Z77 boards all have excellent, solid UEFIs that have been thoroughly bug tested. Functionality.

[chew*]

I think your expectations are a tad high.

Ive looked at bios's in the boards targeted area.
Its a 5 series...that has always been mainstream.

Asus prime x370
Msi carbon x370
Asrock Fatality x4 x370
Biostar x370 the mid grade one.

They are all pretty sparse and dumbed down.

Some lack support for pc3200.
Speaking of which here a real complaint.
B350 boards specced for 3200 vs x370 the high end chipset.

In that category the bios needs to be short sweet and simple and work.

If the top line board comes out like this.....grab the torches and pitchforks then.