Discussion in 'Videocards - AMD Radeon Drivers Section' started by connos, Oct 9, 2019.
Jonas are you positive that 19.7.5 does not have any downclocking issues compared to 19.10.1?
It won't do anything, unfortunately.
I assume that the card does what it wants more so then wattman/AB being able to control it?
I dont undestand. So you are playing Metro for instance and gpu downclocks to 300mhz while in game? How can that be? Isnt that mean the card is broken?
It goes down for brief moments then goes back up to 2000Mhz+ for example.Causing stutter.
However, what's more interesting is that the osd shows that that gpu usage drops as well.
So the card thinks there is no "load" when in fact there is. That's the broken part of it IMO.
Another example is that it can hover around 1100Mhz at the main menu then go up to 2000Mhz+ while in game. But in another game it can be at the correct game boost speeds at main menu and in game.
The way that the card boost vs how it's suppose to boost is indeed broken.
I am keeping any eye out for any reviews that has a store bought card to see if the behavior is actually found vs press release cards given to reviewers. Because this is something I've seen no reviewer mention.
Definitely something to keep an eye on. You guys are just scaring me as a potential future Navi user with these issues. I didnt see any similiar behaviour with previous amd gpus. Dropping clocks ingame menu occurs even on Polaris but it goes up ingame and never downclocks if you feed it with enough power.
Im sure this is a something amd will look into it with future driver updates.
Ah so that's how it sounded like but no the earlier drivers have exactly the same situation with the boost clocks so that's not a new behavior in later drivers.
However the stability and display corruption issues and the random black screens while recoverable and the event log reports are no more but the other driver bugs and the bugs fixed in later drivers (Though some of them only partially.) would still exist on 19.7.5 though comparably this driver has been the best one so far and then for the weekend a few tests on 19.8.2 and 19.8.1 although I expect the display corruption and periodic black screen errors to come back as that looks to be where it started and then 19.9.1 to 19.9.3 never really fixed anything far as these go for users who actually get them at all (Making it harder to track down.) and 19.10.1fixes a few things but can also be more unstable. (That's the one generating the event log kernel errors and media playback via GPU is completely broken citing a issue with one of the driver .dll files.)
Fine tuning the clock speeds and undervolting and I'm sitting around 1800 Mhz out of the specified 1900 Mhz much of the time but any one thing disrupting this and it's back to 1400 - 1600 Mhz plus the Pulse here has a thing with bios and fan settings and not wanting to go over 60 degrees C junction so if that is hit regardless of curve the fans will spin up a bit though that also ensures thermals won't be too big of a bother although noise might.
Leakage and other problems also means it starts smooth and then as RAM and VRAM fill up, CPU's hitting some snags and other hitches and stutter issues and GPU activity just drops but yeah game optimization and problems so that's not unexpected but Navi here gets hit a bit worse since it's so reactive or what to call it. (Division 2 is holding up well outside of player active areas and then Breakpoint lasts about an hour before it becomes more erratic but that also shows other game issues like unresponsive controls so it needs a patch or several. )
EDIT: Opinion on the card itself hasn't changed much but the driver situation is more like 6 months or later instead of a few months for getting the early driver issues fixed up and it's just a few more weeks until month 5 I believe and this and a few other issues and regressions do need fixing whatever is causing them.
This with the clock speeds seem just like overly aggressive adjusting to however many different factors there are for the boost parameter whether up and down and then for the rest well since it's not affecting everyone there's too many factors to really say anything.
(Older DP or HDMI versions? Win10 builds? Older hardware in general? Certain software and all kinds of things.)
EDIT: Whether intended or not, seems like it should have some delay before readjusting but the D3D9 and 1080p issues also indicate a problem with detection of how much the GPU is doing or something since it's just dropping to almost idle speeds in these conditions which should be impossible as any GPU load at all should avoid those levels.
(And from posts so far GPU load can still be near or around 50% so it shouldn't hit idle speeds or anywhere near but still does.)
Meaning there are probably some detection bugs or something of that nature as well and not just the GPU throttling down any chance it can because that's some tuned parameter or behavior AMD is using for whatever reason unless it's at or near 100% load which isn't entirely feasible as any number of conditions or little hiccups can drop it down at least a bit.
(And then the GPU lowers speeds by a fair amount almost immediately.)
EDIT: And yeah it's mostly about downclocking here in topic though the GPU can also spike above the set boost frequency though this can be mitigated by the power slider in Wattman either keeping it around 0 to 10% or thereabout or negative by limiting power draw though depending on GPU model and any bios values for this the total might be a bit variable and it's the GPU core power and not the total or at least that's what I see being affected via GPU Z when comparing.
So with the current settings it's hitting around 150w with the default bios at 195w and the silent bios at 180w if I remember correctly for the 5700XT Pulse so increasing it upwards would near 300w and scaling it down would near 100w if I'm getting these conversions right.
(That's for what it can draw though but having a higher threshold here affects boost maximum and having a lower can reduce it but can also limit the average sustainable clock speeds too if it's lowered a bit below what is needed for these.)
Spikes are very short term though and then go back to normal values and should be just fine as long as it's not sustained over a long period of time.
It downclocks on 19.7.5
I cannot get rid of the stuttering in either Apex Legends or the cinematic videos of the single player campaign of Battlefield V, due to the card aggressively downclocking. In case of BFV, the game itself runs well as far as I can tell, even under DX12.
What I try to do is set all graphics settings to the max and increase resolution scale beyond 100% to put the load on the GPU and prevent the core clock from downclocking. Usually the core clock stays at 2000MHz+ if the readings are correct. BF1, BFV, Titanfall 2 and FH4 all run well like this and don't downclock much. But these are all well optimized games. You just have to find the right balance between maxing out the graphics to prevent the downclock without murdering your FPS. In my case, preferably not less than 100FPS. On the other hand, Insurgency: Sandstorm is not that great performance wise. Maxing out the graphics with and without increasing resolution scale, reduces the FPS significantly and sometimes it goes below 60. Some maps are better than others. On the other hand, lowering the settings becomes a stuttering fest with the downclocking. Wish it was as smooth as the others. Would be more enjoyable actually.
Edit: Found out that shadows are causing the inconsistent framereate and FPS drops in Insurgency. Unfortunately, setting the Shadows to low makes the game look ugly.
Does this actually work? If yes, where can I find a working BIOS for the 5700xt reference?
From here https://www.techpowerup.com/vgabios/
Those are a lot to chose from which brings me to the next question: Which one to chose for a reference card? I don´t care about adjusted fan curves or overclock. I just need a firmware which hopefully won´t clock down that much. Thank you very much, really appreciate your help
Doubt the bios would do anything for that, you get different settings for the clock speed and voltage but it can all be set in Wattman and doesn't affect how the GPU is clocking down during less demanding situations but something's off with at least a few games where this is done too aggressively and for possibly the majority of D3D9 titles including some newer popular titles.
Reducing throttling by tweaking voltages and going with a soft power play table mod can help with hitting higher clock speeds but it doesn't affect how the card boosts unless it throttles specifically from power or thermal limitations which isn't the problem here it's something in the driver behavior.
Stock bios on the 5700XT and in particular the 50th AE editions should be the highest clocked in fact and several custom models just reduce or balance these values with the overall effect of most cards hitting within the same performance range and it's mainly a difference in temperature, noise and power though the higher models do have 2x 8-pin instead of 1x 6 and 1x 8 pin but the power draw isn't really that decisive for GPU scaling and can even be detrimental in my own testing at least though water cooling might be a way but we're pushing into 2100 Mhz for that to be a factor then.
Minimum doesn't seem to do much either and the card operates a bit as it wants often boosting to around 50 - 100 Mhz below the target setting in Wattman or the bios defaults.
Voltage of 1.2v isn't actually terrible but lowering it down allows the card to do the same or that's the idea and with a temperature reduction of 10 - 20 degrees Celsius for the junction temperature reading meaning less possible throttling.
SOC is also a thing but most AMD models use 1.2v here too and most custom models appear to set a bios value of 1.05v with no real negatives other than a slightly reduced temperature from the other chips on the board.
Effectively the driver needs work and other than that the GPU is doing as it should operating around 1900 - 1950 Mhz if you are targeting 2000 short of if it's throttled by heat or power but even the stock models can do well only with a slightly higher fan curve and a bit added noise as a result so it's the driver code that hampers things and that's not something that can be forced. Unfortunately as that's one of the remaining bigger issues with this card for the titles affected by this as it completely messes up performance but also smoothness or stability or what to call it by hitting stuttering and sudden hitches or stalls as the GPU clocks keep re-adjusting and falling to very low values even more than the card should be doing for a lower demanding title.
EDIT: Doesn't mean it can't be tried or anything but it seems more like a way to get a default tweak going without having to rely on Wattman and it's little quirks and limits but not entirely due to driver quirks affecting also OverDrive itself.
Balancing the GPU is going to potentially reduce it hitting power and heat throttling points and might make it a bit more silent or allow for a higher boost clock and maintaining it consistently but I don't think you can hope for much for this downclocking problem until the driver code is properly fixed whatever is really causing it during lower GPU demands and whatever is going on with D3D9 too.
Plus possible issues with keeping the card operating at full efficiency and not start reducing clocks for any little thing or hiccup, I can't really say and I don't have the knowledge or skill but I feel a sort of restriction before it starts reducing speeds instead of immediately throttling wouldn't be a bad advanced setting for tuning the GPU behavior but it might not be possible to have some timeout or few milliseconds delay before this behavior kicks in for how the card just operates and works in hardware.
(Or how to say, it's designed to freely boost and clock up and down instead of use the P state system in earlier GPU models after all even boosting above the set limits for short bursts of time occasionally.)
What card do you have? There is no way to avoid the downclocks. For this we are waiting a fix from AMD. But some people said it help with the bios. I have reference and I can load the Sapphire nitro+ settings.
Here is a reference bios https://www.techpowerup.com/vgabios/212169/sapphire-rx5700xt-8192-190616
There is only one diference between the bioses. The GPU Amp power limit. For example with reference its 170 A but with nitro is 189 A, Red Devil is 196 A. Some people said it helped with stuttering but downclocking is still there.
I also have an 5700XT reference card. I dont really care about the downlocks as long as the stutter is gone. I am totally fine with the GPU running at 300mhz if the games do not stutter. So from what I understand from your text is that you are running the Nitro+ Bios and your card doesn´t have those stutter issues. Will give it a go then later today.
aren't these user rectifiable by turning up the game settings
Raising the power limit should scale with the value here allowing a +- 50% although from testing the card kinda tanks when going negative so from 0% to -10% and I'm sitting at 1600 Mhz core clock speeds already down from the targeted 1900 Mhz for this test.
On the other end of this increasing the value doesn't do much either unless you are pushing the core clocks into the 2 Ghz range so from +10% and I'm sitting around 1850 and above that it doesn't do anything though it does draw more power and increases the temperature by a small amount.
(Was hoping it would be a limit for how much the GPU can draw but it's more that it draws this amount of power and can use that to scale if required rather than increasing as needed by having a higher limit to where it caps.)
Think the silent bios of the Pulse here has a target of 180w and the default targets 195w and then Nitro and some others go up to 210w plus some variances in target temperature (105, 107 or 110 degrees core junction temp.) and the fan profile though this has been wonky since 19.7.2 at least and it frequently overshoots or starts increasing speeds even during low activity compared to 19.7.1 or 19.7.2 (The first two Navi GPU drivers.) so some tuning is required to get this optimal as a result. (Noise against cooling efficiency.)
For extreme overclocking of sorts it looks like a +50% power and water cooling can allow up to 2.2 Ghz core clock but for benchmarking mostly and stability is a bit so-so during general usage plus it fluctuates a bit for what it reaches when targeting a higher value like that. Same for memory when pushed above 950Mhz though it also gives a nice performance boost if it can keep stable at 920 or higher compared to the stock 875 although the modules are also pretty sensitive.
Yep the general workaround is to either go with VSR or try and set higher settings to force the GPU to operate at or near 100% and thus keep the clock speeds near the boost limit though when it tries to drop it can also cause stability problems although later drivers seem a bit better at working like this by just trying to force it to run at full load.
EDIT: The GPU hitting a few drops though maintaining higher clock speeds can also lead to more stuttering and framerate and frame time variances and fluctuation though it also varies for some games being more affected than others and it might also be a CPU limit of sorts also affecting the GPU performance since some titles like CS:GO are CPU heavy while relatively GPU light though then again when the CPU is the bottleneck then that's going to affect performance more than trying to tune how the GPU performs until it can be a bottleneck such as running with downsampling from a higher resolution.
Not the best of explanations.
Destiny 2, Borderlands 3, DOTA and CS:GO for a few affected titles and the first two should be using D3D11 at least so not just from whatever is affecting how the card handles under D3D9 plus issues with GPU driver crashes being frequent although I believe the newer drivers has made some improvements.
Still much I don't completely understand here on how and why this happens and what causes it to drop and when it does why it also affects stability.
One thing is from undervolting too hard though but even without that it's still not completely stable going from posts by users who haven't even accessed Wattman.
(Could be any number of things though, complicated code and all sorts of possibilities.)
EDIT: Though that's not changing this problem and anything that's less demanding seemingly causes the card to lower the core clock speeds down which for D3D9 can result in very low values instead of keeping around the mid range of the clock speed range so ~1500 Mhz or so at the lowest.
Whereas for some situations it almost goes down to zero or close to hitting around 30 Mhz or so even if it's brief that still causes some pretty significant stuttering and such.
That seems to be the key difference with D3D9 at least where it drops almost down to inactive or idle speeds instead of just falling to the mid point or "game" clock speeds instead of the "boost" full clock speed value.
(Of sorts, generally it'll be within 50 Mhz below or above this target value.)