Discussion in 'Videocards - AMD Radeon' started by Glidefan, Mar 31, 2019.
I've been using a Vega 64 for a while now, the Asus ROG Strix one, and off and on been trying to tweak it a little, but only recently began to undervolt it, inspired by this topic here!
So my HBM seems to go to at least 1140MHz with 990mV, at least it gets through Firestrike Ultra just fine (of course stability tests will be commenced later).
However, when I try to undervolt the P7 GPU clock state to anything lower than 1150mV, I get a driver crash after a few seconds of benchmarking. I'm able to raise the P7 speed to 1700MHz on 1200mV (maybe even further, but I didn't test it yet), and perhaps a bit lower, but stuff like 1030mV like a lot of you guys doesn't seem like a possibility.
Raising or lowering the Power limit between 0% and +50% doesn't seem to make a difference in stability too (though a higher limit improved FPS (less throttling?), but also raises temps quicker).
I'm still on 19.5.2 drivers, but I don't expect newer ones to improve it by much.
Any suggestions? Or did I loose the silicon lottery...
Maybe the LC bios will do you good, You get a bit more power out of it and Voltage I think
Power = Current * Voltage
As power usage goes up, heat and current interaction leads to a positive correlation with higher voltage requirement.
Given that the Asus ROG Strix Vega 64 is deigned to run at 260 watts, it may not be able to run as low voltage levels as a card designed to run at 220 watts. Of course, significant reductions in temperature may allow lowered voltage operation, but there may still be fundamental differences in design necessitating the original higher power requirement.
That said, the complexity of the relation necessitates testing individual cards to ascertain changes in behaviour at different operating states.
Will it matter which liquid BIOS I would flash to?
Before You get to Flash things
make sure You learn Your Vega first, when You do jump into more complicated matters ( LC Flashing).
Been changing some UV/OC settings with latest drivers. Found lower sweet spot UV without losing much(1%) in terms of performance. Also Hot spot runs about 7-8c cooler. Played Division2 for 45 min no crash as well as FSU stress test(98.9%) looping Heaven a few hours will be true test.
P7-1712mhz @ 1150mv
P6-1607mhz @ 1100mv
P5-1462mhz @ 1050mv
P4-1202mhz @ 1025mv
P3-1187mhz @ 1000mv
P2-1132mhz @ 950mv
HBM-1115mhz @ 1050mv
Division2 avg low Clk- 1635-1637mhz
Max watts- 305w @ 1.131v
2560x1440 Division2 - 86 FPS
Default-Timespy- 7914@ 54.91fps/43.08fps
Default-FireStrike- 25854@ 127.35fps/100.61fps
P7-1697mhz @ 1096mv
P6-1537mhz @ 1050mv
P5-1402mhz @ 1050mv
P4-1202mhz @ 1045mv
P3-1137mhz @ 1000mv
P2-1082mhz @ 950mv
HBM-1115mhz @ 1050mv
Division2 avg low Clk- 1630-1640mhz
Max watts- 277w @ 1.087v
2560x1440 Division2 - 85 FPS
Default-Timespy- 7826@ 54.28fps/42.63fps
Default-FireStrike- 25552@ 126.45fps/98.86fps
I am trying to learn the card, but it seems like an odd beast to tackle.
I can overclock and undervolt the HBM just fine (1140MHz, 975mV at Power Limit+25%, however, by increasing the PL, the HBM-voltage needs to be increased too, but it's mostly for benchmarking). I need to test the difference between Memory Timings 1 and 2 yet.
But the GPU clock is a drama. At +25%, the P7 set anything above 1680MHz causes a driver crash (how should I test for P1-6 settings? Work my way down from P6?).
And I can't lower the Voltage below 1160mV at P7 apparently without crashing.
So it probably needs more testing and tuning...
With the 'working' settings the GPU clock speed averages around 1480MHz in FireStrike Ultra, and I get a score of just under 5500 points. I don't think it's particularly good or anything. The temps get just over 61 degrees.
P1 to P5 is best to have on default with UV (or Tweaked some more towards Efficiency)
rest is ok.
Im mostly tweaking only P6-P7
Keep PL at 0%, HBM freq at stock (945) and volt 951. P1 900v, P2 950v, P3-P6 1000v. Now start lowering P7 voltage. Mine works at 1060v. Lower then that and sometimes crashes, sometimes works. 1060 is 100% stable. When you find the lowest stable voltage at P7 start OC HBM.
Edit: About PL. P7 at 1632 and 1060 i don't need more then 0%. Never. When you find a stable P7 UV then start raising PL by 2-3% at a time until you find the higher you need. Don't go over it cause some times wattage spikes a lot and that's when you get a crash. It would be good to make profiles for your games with different PL settings to avoid unnecessary spikes
What I've noticed about PL the most is that it allows the card to reach higher clocks as well as not fluctuate as much in game. I did some testing in Division2 to see diiference while I was doing last round of testing heres what I was seeing:
PL 0%- 1452-1481mhz @ 242w
PL 10%- 1529-1556mhz @ 265w
PL 20%- 1593-1600mhz @ 288w
PL 25%- 1619-1630mhz @ 299w
PL 30%- 1635-1637mhz @ 305w
PL 50%- 1635-1637mhz @ 307w
Now those clock speeds and wattages will vary depending on you settings, but it does give you a window into whats going on. I've tried every youtube, forum setting and for my card I've never had luck with leaving PL at 0-20% at least if we are talking about pushing for 1630+mhz in game. Thats not saying you wont. But from my experience PL is needed. I have the PL set at 50% on my setting from post #277 and it ran Heaven(1440p Maxed) looped for over 9 hours stable with a 245-250w avg. What you have to remember to is that your clocks, wattage, etc will vary depending on the game. I've been using Division2 alot because it seems to push the card hard. Example: 277w in Division vs 250w in Heaven. Other games vary also. So its best to find a game or stress test that push the card for stability.
I tried to learn this Vega64 and I've failed maybe ... the only thing that is stable and has sustained higher clocks for me is:
P7 1677 MHz (set) and it actually is: 1615 - 1650 rarely goes to above 1660) @ 1150 mV PL: 50%
I've tried several other things and ideas from colleagues from here but either it crashed either it had lower clocks. So if I want to go above 1600 MHz core all the time and hoover between 1625-1650 MHz I need PL to be 50% and volts 1150 mV.
Regarding HBM it's only stable at 1050-1750 MHz and voltage doesn't seem to affect it or change it. its default at 1100 mV sometimes I've seen 1350 mV but not sure how and why.
Bottom line for me:
P7 1677 MHz @ 1150mV and PL:50% HBM 1000 MHz @ 1100mV. This is the best I can do maybe someone could squeeze a little more but I can't
The power consumption in watts the max I've seen was around 380W in Wattman OSD in "A Plague Tale Innocence" (pc game) at 2.7K resolution all maxed out. I think PL was soft modded to 75% or 100% and the card is still alive.
P.S. 1688MHz is not doable at 1150mV for me ... maybe PL is too high and spikes kill it but fore a few MHz I prefer stability and much more stable clocks that PL 50% - 75% offers.
It quite a good card when it comes to performance I think is a little above 1080GTX and bellow TI.
There are a few games optimized for RYZEN/RADEON that actually work really nice. I've got it for a Hackintosh and for 2.7K gaming and if you can live with 1-2 settings high from ultra here and there it will peform very nice. The refresh rate that I play on is 75Hz freesync and I'm at 73 fps 99% of the times in all games.
I've done some testing with mine and these are what's stable for me so far (after several stress tests passed):
Power Limit: +25%
-For me this appears to be the option which gives the best temperature-performance ratio. A lower limit gives better temps but seems to hit performance by quite a bit. A higher limit increases performance, but not by much, while it still gets hotter.
HBM2: 1100MHz @ 975mV
-I can succesfully complete 3D Mark Firestrike Ultra benchmark runs with speeds up to 1150MHz without artifacts appearing, but the stress tests fail after quite a number of loops. 1100MHz passes those at least. 975mV seems to be close to the threshold: a bit lower starts impacting performance, but higher doesn't improve performance, while it may raise temperatures.
P1: 992MHz @ 900mV
P2: 1082MHz @ 950mV
P3: 1197MHz @ 980mV
P4: 1337MHz @ 1000mV
P5: 1472MHz @ 1030mV
P6: 1592MHz @ 1060mV
P7: 1662MHz @ 1170mV
I overclocked P4-P6 by a bit, and adjusted P1-P3 just by moving their respective knobs in the Wattman graph.
Overclocking P4-P6 seemed to boost performance in Firestrike considerably and I undervolted bit by bit (just by running the first FS graphics test. If it fails after two seconds I knew the settings should be loosened up a bit. Failing the resulting Stress Tests made me loosen the voltage settings slightly until that one completes just fine).
As I've mentioned before, my GPU doesn't really like ondervolting or overclocking too much. I get a crash when P7 is set to clocks higher than 1670, or when I try to undervolt below 1160mV.
I've also set a more aggressive fan-curve, especially with the current weather conditions. It's a bit more noisy, but not as much as AMD's reference board fans.
I'm going to play some Battlefield 1 with these settings and will keep you updated.
EDIT: Battlefield 1 didn't like the P1-P6 overclocks. So I've set them to default while retaining the Undervolt settings, and the game worked fine.
God vegas 14nm are geting faster and faster, the scaling on R7 is strange.
wow that 290X just refuses to die. What a beast of card that is.
Yes and V64 LC is "only" at 1668MHz boost (Default Balanced mode) (But You have Full 64CUs)
BTW 1750Mhz Boost is only avaible through Manual OC or Turbo Mode in WattMan
Yes 1650Mhz is enough for V64....
So the Radeon cards are being tested on "defaults" as opposed what people would actually play them with? Bogus testing!
I can't say I'm not happy but why are they keep getting faster ?
All the R7 are is Vega tweaked on 7nm. Sure they have higher clock but like OnnA said V64's have 64CU vs 60 on R7's so its levels the field alittle. Now if the R7 had a full 64CU plus clk increase they would get around 5-8% more FPS. They would be right there with 2080/1080Ti's on every game/bench.
R7s are what AMD envisoned 2yrs ago when they released Vega. They just didnt refine it enough at the time. I feel like the are about one product cycle behind nVidia. The new RDNA architecture should have came out when Green team released Turing. Hopefully they can catch nvidia with there pants down and get caught up alittle. I feel like this "super" card crap is because they dont have anything ready at the moment.