Discussion in 'Videocards - AMD Radeon' started by OnnA, Oct 29, 2020.
Wanna raise a thread on AMD forums? I will submit a bug report, others should do the same.
Thanks for the info! Ran CPU, no errors. GPU, no errors. Memory, no errors, VRAM no errors. The Power Test though... A handful of errors occurred in two minutes:
For how long? I mean the various tests?. You need to do some research about these errors, PSU test should have resulted in PC reset/abrupt shutdown on load for a bad PSU. Try to see the log values for 12v, 5v and 3.3 volts during PSU test, the should not deviate much from their normal values much.
If it's not too conservative then after a while on auto / default settings it might be able to hit the 3300 value, Wattman overall feels like it's conservative so you have a slight offset for the set value and what the real figure actually is.
-50 Mhz on the GPU core, -10 Mhz on the VRAM for clocks.
Not sure about voltage and power draw actual values and the fan speed feels like it can be whatever so I usually set a short range around 62 - 70 for these and then during load it's at least somewhere around where it should be.
6000 series changes the fan speed for idle again because that also goes back and forth so it's up to user preference and tolerance the customs likely have a big advantage pushing more air and better resulting temps at a lower noise level.
Bugs do exist though a number of users can't get the voltage to take and it goes back to bios defaults instead it's usually the max supported anyway so not a huge deal but for undervolting fine tuning it causes some problems.
Fan profile behavior on older GPU models and how it changes for Navi10 has also gone a bit back and forth, overall it just needs a more thorough overhaul or code checking making sure it works as intended and maybe also correcting this weird offset where it can't ever hit the max number it's always a bit lower.
Conservative meaning that the fan profile keeps close to minimum speeds even the zero fan mode and then once the thermal limit is reached it boosts up near max until the GPU is below this threshold before winding down.
I like using a manual curve/stepping whatever it is now to avoid these bursts instead keeping a constant higher speed.
Avoids the GPU fan behavior entirely and while it gets load at 60+ percentage or so it never gets really loud at hitting 70 to 80% speeds for the reference cooler.
Doesn't go on max speeds even on Auto and temp over 110 C (for hotspot).
One more thing I wanted to ask you, I suffered a PC hard freeze with buzzing sound during a not so demanding game "Black Squad". Haven't been able to reproduce it through multiple runs of 3d mark, Port royal stress test, 2-3 hours of cyberpunk and even the game Black Squad itself. All OCCT tests produced no errors as well. I know there could be many reasons for it but what do you think this could be based on your experience? Updated Audio drivers and reinstalled GPU drivers as well since then.
Yeah it doesn't have to do anything demanding it just has to do a incorrect instruction and that can cause the drivers to act up really badly.
AMD worked out a lot of that with Navi10 and the 2019 and 2020 drivers but DirectX 9 is still problematic and DirectX 11 isn't quite performing as well as it should.
Unreal Engine 3 has a variety of flaws as well but usually it doesn't lock up the system and the most you get is a software crash with the game itself at some times also a extra as the WDDM and display driver crashes and resets.
I still primarily use DXVK on everything not natively DirectX 12 or Vulkan even if Navi20 has some improvements because it has similar flaws in the drivers, overall even going via wrapper Vulkan just works a lot cleaner in my testing.
(Stabler, performs better and the driver scales better so GPU utilization is a lot smoother as is frame timing and once the shader data is cached it's a great experience overall.)
32-bit and D3D9 looking at the specifics of this game it should be easy but AMD's taking their time working out the issues with this API and the RDNA instruction set. (Navi10 and Navi20 GPU's)
Not all D3D9 games are broken either, some just under perform like Witcher 2 until a late Autumn 2020.x driver fixed it.
The November driver I think also fixed one of the core bugs with broken shader model 3.0 vertex instructions for D3D9 but it's not fully fixed up yet for support and reliability with this API and the new GPU instruction set of RDNA.
Performance at least sounds like it's eventually going to be fixed up, reliability and stability though well hopefully that too and less specific game profiles and overrides.
If that's what this is caused by and nothing else I would just hope that newer drivers can fix this finally, otherwise there's DXVK and wrapping but online and MMO games don't always take well to this or other software due to a variety of anti cheat measures that might block the interaction entirely or lead to problems including bans at worst.
This bit from Buildzoid could also be part of the problem and the answer to at least some of the instabilities going on when the GPU's working at a lower workload.
That would cause a pretty bad crash and the way it works you have a curve so you only specify the maximum voltage value at the highest clock speed instead of setting a minimum voltage.
If gets to the point it uses idle voltage numbers then that can't really be set at all it tends to mostly just use the bios values.
With a slight exception for high refresh rate displays pushing memory speeds to their active state and using a bit more power as a result.
So currently I have my Merc 319 6800 XT Game Boosting to 2454 at 30% fan hitting 70C in game. I'm wondering what this thing can do at 100%. Going to try to crank them up.
Just hit 2600 core stable with 50% fan speed on this thing! I had no idea this had it in it...STABLE! Power limit +15% of course. Did not bump the memory yet. Yes, using AMD's Tuning software.
Hey there Ghost XL,
I have the same MERC 6800XT and it is awesome, although I had some crashing with some games, after I reinstalled Win10 all became a lot better(Old Vega64 leftovers I guess).
Even so, in OCCT Vram test after 1 hour I have 3 WHEA errors average, can you check on your card please? Want to know if it is software thing, or if I should RMA my card... OCCT full system test got no errors, only 95% VRAM did.
Really appreciate if you can check on your side, and every 6000 series could pitch in if they want! thanks
95% VRAM threw 4 of those errors on me, but stayed that way, and they threw those errors very early within 5 mins. I don't think this is anything to worry about. It has been rock solid despite throwing those. I have the VRAM at default clocks, core 2600 at this time with the +15% power increase.
IMHO I would not worry and enjoy the card.
GhostXL dude you rock,
thank you very much for the reply, it makes me think that these erros are a more normal thing. Because it made me question every crash from my 800€+ card before with fear of defect
Again thanks for the effort, and its moments like these that makes me appreciate this awesome community
Ah so the voltage going back to stock is a intended behavior it seems.
If I am reading this right then in Wattman adjusting the value lets the GPU scale to the voltage value but this seems to only affect the regular clock scale and once the GPU boosts past this it can go back to a default value so 1.015 for example for the 6800 non-XT.
@OnnA That basically confirms yet again what I have said.... 6800xt on par or slightly better in RT than RTX3070 (+- 10%)... And on top -- 6900xt a little bit behind 3080 (13% more or less).
However if you combine it with raster performance then the AMD cards are really beastly in performance and UV and OC ten times better than NV counterparts.
Newest Hardware Unboxed SAM benchmarks:
Brings a nice boost in a few games, in few a minor boost (3-5%) and in - some cases - a few percent of negative scaling....
The biggest improvements in newest games/engines shows that it will only improve with once the Developers will update their game engines and optimize for resizable BAR.
BTW.... Based on the current situation I am actually happy that I have payed only ca 80-90 EUR on top of the MSRP for the 6900XT... based on the recent information and feedback regarding the whole mining craze going on and the things ASUS is doing with the "US market" (basically +80-90 USD for MoBos and GPUs "per se") the prices will well double for ANYTHING usable for mining.....
You can expect virtually NO STOCK for both AMD and NV for the foreseeable future....
You can make Cryptocurrencies worthless, that would solve that problem.
It's not that easy @Chastity .... The crypto is worth as much as others are willing to pay for it....
And making it digital and harder to trace makes the whole thing... not necessarily easier for all of us.... and more interesting for others in the grey zone....
And yes... I dug cryptos when Bitcoin started... Had over 12 coins back then... but as a poor student I have sold them for .... 37 PLN each... translates now to roughly 7 Euros.... But I have had no money back then... now for those 12 coins I could buy a mid-end flat... ohwell ...
No problem. I had errors thrown in most cards I've owned...but usually below double digits. As long as you don't go crazy on the Memory OC I'm sure it will be fine. I think this series benefits more from Core OC as it is anyway.
The VRAM clocks and other restrictions limit the gains and the core 256-bit bandwidth is likely also a hindrance especially 3840x21560 and 5120x2880 or higher pretty sure the infinity cache would struggle though the GPU itself isn't really 5k capable though the 6800 non-XT is surprisingly capable for mid-framerate 4k gameplay in my own testing.
Or high-refresh really if you're using a 60 to 75Hz panel, SpecialK a few tweaks and limiting to 72 FPS is how I do it at the moment some dips but it's holding up well.
Curious what these could really do if not bios and driver locked the way they are though.
Would think 1.150v maybe up to 1.200v could have let the GPU's hit closer to 3 Ghz but you have locks on GPU clock speeds, voltage, power draw and the various timing restrictions and what else.
Even with that though you can almost bridge the gap to the 6800 XT though once that boosts it takes the lead again and then the limits prevent the 6800 XT from hitting quite up there with the 6900 XT but it also gets really close.
Would expect the limits to exist for that reason in particular, high clocked custom 6800 with a 2.5 to triple slot cooler or water with a solid PCB for power delivery and with enough wattage and GPU voltage the extra cores will be less of a factor than the resulting clock speed gains although still overall limited by the memory bandwidth and bus width once the infinity cache memory is starting to hit more misses / limitations for 4k and higher.
(The Navi21 H version sounds really interesting if that's HBM and what this can do because there might still be a memory infinity cache limit of 128 MB but the 256-bit bus is removed by HBM stacks up to 1024-bit and how this works with stacks and such.)
EDIT: So where was this going oh yeah I think VRAM can show potential gains but the timing and corresponding clock speeds or ignoring the timing and going for higher clock speeds but still hitting limits makes non-benchmark gains fairly low.
5 - 6% increase and ECC could be kicking in limiting gains so it's a slight 2050 - 2100 Mhz range (Double data rate.) without being able to adjust memory voltage or loosen the timings to clock it higher without error correcting lowering performance or hitting artifacts or instability.
Several early reviews seem to just increase that to 2150 without fully testing it though which could work but might already be limiting things if ECC's going in, 2140 effective because the settings differ a bit with actual effective value and then a bit lower because for some reason too the manual value makes a different final value over the percentage slider in normal mode.
Variance from every GPU too so there are several cards that can take the full maximum 2150 setting too it's not like it's impossible or anything it's less sensitive than the Navi10 cards were about memory clocks and where ECC kicked in.
EDIT: Oh and also what the effective max safe voltage would be, 1.150 to 1.175 perhaps but it could be lower than the 1.2v on Navi10 with this process.
Kinda locked so kinda hard to find out, plus how it affects a offset or curve when adjusted so it's not really limiting the voltage it's just scaling the GPU boost behavior and can still hit the default / max target even if the GPU tries to scale to around the set lowered value.
Need to do some more testing with this myself after reading that undervolt and underclock article from Igors Lab on the 6000 series GPU behavior and how this worked.
Jonas --- safe voltage for 6900xt (or rather --- default) is 1.175 so it should be considered definitely as "safe" for the navi21 derivates. But yeah... this time Igor did a really nice job on analyzing the RDNA2 behavior and how it reacts to UC/OC/UV/OV.