Vega owners club

[turbogear]

Guru3d put up a review of the ASUS Radeon ROG RX Vega 64 STRIX:
http://www.guru3d.com/articles_pages/asus_radeon_rog_rx_vega_64_strix_8gb_review,1.html

From what I read there STRIX does not sound very interesting solution.
I think custom liquid cooled reference card is much better than ASUS's solution.
It does not seem to bring much better performance than the reference card.


Editted:
Interesting enough Guru3d has taken down the review that they did on ASUS Radeon ROG RX Vega 64 STRIX:
http://www.guru3d.com/news-story/review-asus-radeon-rog-rx-vega-64-strix-8gb.html

Here is the reason that Hilbert Hagedoorn wrote:

This morning I received a phone call from ASUS, asking us if we’d be willing to take down the article for a few days as they have made a mistake. The sample we received did not get a final BIOS for its final clock frequencies and fan tweaking. Ergo, the sample we received carries a default reference BIOS.
It’s a colossal mistake, but as such the end-results in the review are not representative enough for the final product. ASUS will get the finalized BIOS over once they have finished (likely a day or two) after which we will re-test the card with that final BIOS and thus republish the review. All this explains why the STRIX card was so incredibly close to Vega 64 performance. Apologies for the inconvenience, but this mistake was not one coming from us.

Makes me wonder how ASUS could do such a mistake.

 

[kurosagi01]

So i've bite the bullet and ordered myself a Sapphire Vega64 =)

 

[kurosagi01]

What numbers do people recommend testing first with undervolting and power limit at stock speed as i am new to the whole undervolting thing. Been reading different sites on user experience and all numbers vary since every card are different.

Source *from reddit amd group*:
https://docs.google.com/spreadsheet...beRiOTMsGeIkCDQUhXSBA/htmlview#gid=1964178281

 

[londiste]

do you want increased performance or lower power consumption?
for increased performance, undervolted or not, power limit +50% is a must. and power consumption will increase.
vega does match the clocks to core voltage well and power consumption changes accordingly.

as for actual testing... latest beta version of hwinfo64 should be (somewhat) able to monitor clocks and voltages. in my experience though i was looking at a cheap power meter measuring the whole system power consumption from the wall to see if undervolting helped.

as for the process - put some load on the gpu. steady heavy load is best. then set the voltages in wattman (its buggy so from all sources it looks like you need to change the hbm voltage to really change gpu voltage so i ended up changing all three at once), going down by some increments like 20mv until you see issues or voltages-power consupmtion no longer drops. for me power consumption no longer dropped from 950mv lower (in wattman, it was really probably a little over 1000mv).

keep a keen eye on the clocks though as these will start dropping under load when you decrease voltage.

 

[kurosagi01]

do you want increased performance or lower power consumption?
for increased performance, undervolted or not, power limit +50% is a must. and power consumption will increase.
vega does match the clocks to core voltage well and power consumption changes accordingly.

as for actual testing... latest beta version of hwinfo64 should be (somewhat) able to monitor clocks and voltages. in my experience though i was looking at a cheap power meter measuring the whole system power consumption from the wall to see if undervolting helped.

as for the process - put some load on the gpu. steady heavy load is best. then set the voltages in wattman (its buggy so from all sources it looks like you need to change the hbm voltage to really change gpu voltage so i ended up changing all three at once), going down by some increments like 20mv until you see issues or voltages-power consupmtion no longer drops. for me power consumption no longer dropped from 950mv lower (in wattman, it was really probably a little over 1000mv).

keep a keen eye on the clocks though as these will start dropping under load when you decrease voltage.

I'm happy to run the card at stock performance but lower the power consumption as i'm not big fan of overclocking and if i did probably a little boost.
I'll get installing hwinfo64 when i'm home and for work load i've seen people use heaven benchmark running in background. How long should i run the test for to know the voltage is at safe level before bringing it down again?

 

[turbogear]

I'm happy to run the card at stock performance but lower the power consumption as i'm not big fan of overclocking and if i did probably a little boost.
I'll get installing hwinfo64 when i'm home and for work load i've seen people use heaven benchmark running in background. How long should i run the test for to know the voltage is at safe level before bringing it down again?

As mentioned in post #34 in this thread, I am using Fire Strike Ultra stress test but unfortunately this benchmarking tool is not for free.
It is part of 3DMark which cost 29,99 dollars.
For unstable voltages, I uaually see a system/GPU freeze with this stress test.
I used Unigine Heaven before, it did not stress the GPU sometimes that much. Some voltages which ran stable in Heaven crached in Fire Strike Ultra.

Furthermore, I tested by playing Witcher 3 which also puts quite a stress on the Vega as I mentioned in post #32.

I did not also want to overclock the Vega to the limits. I justed tried to find best case scenario with higher performance but not too higher power draw penalty as explained in post #32 and #39.

 

[acperience7]

What numbers do people recommend testing first with undervolting and power limit at stock speed as i am new to the whole undervolting thing. Been reading different sites on user experience and all numbers vary since every card are different.

Source *from reddit amd group*:
https://docs.google.com/spreadsheet...beRiOTMsGeIkCDQUhXSBA/htmlview#gid=1964178281

My card is a bit faster than stock, but I'm running 1577 core @ 1070mV and 990 memory @ 1050. This is with a 10% power limit increase. My fan is set to 600 (idle) and 2700 (max). Temp targets are stock (75 target and 85 max). I haven't seen 80 so far. Usually hovers between 77-71 depending on what I'm playing. Power usage at the wall is just about stock. The biggest thing for me is that my clocks never dip below 1500 with these settings.

If I were you, I'd start with 1100mV and 15xxMhz work my way down on voltage. That's what I did. However, be aware that the unless you mess with the power limit the card will more aggressively throttle the clocks. I couldn't keep my clocks above 1500 without increasing the power limit (at a lower voltage that is).

My Firestrike graphics score went from 21,5xx to 23,0xx.

 

[kurosagi01]

As mentioned in post #34 in this thread, I am using Fire Strike Ultra stress test but unfortunately this benchmarking tool is not for free.
It is part of 3DMark which cost 29,99 dollars.
For unstable voltages, I uaually see a system/GPU freeze with this stress test.
I used Unigine Heaven before, it did not stress the GPU sometimes that much. Some voltages which ran stable in Heaven crached in Fire Strike Ultra.

Furthermore, I tested by playing Witcher 3 which also puts quite a stress on the Vega as I mentioned in post #32.

I did not also want to overclock the Vega to the limits. I justed tried to find best case scenario with higher performance but not too higher power draw penalty as explained in post #32 and #39.

£22.99 on steam for 3dmark so i'll probably buy it and run firestrike myself aswell for stress test.

 

[turbogear]

£22.99 on steam for 3dmark so i'll probably buy it and run firestrike myself aswell for stress test.

I have also the steam version. I paid 27,99€ over there.

 

[kurosagi01]

My card is a bit faster than stock, but I'm running 1577 core @ 1070mV and 990 memory @ 1050. This is with a 10% power limit increase. My fan is set to 600 (idle) and 2700 (max). Temp targets are stock (75 target and 85 max). I haven't seen 80 so far. Usually hovers between 77-71 depending on what I'm playing. Power usage at the wall is just about stock. The biggest thing for me is that my clocks never dip below 1500 with these settings.

If I were you, I'd start with 1100mV and 15xxMhz work my way down on voltage. That's what I did. However, be aware that the unless you mess with the power limit the card will more aggressively throttle the clocks. I couldn't keep my clocks above 1500 without increasing the power limit (at a lower voltage that is).

My Firestrike graphics score went from 21,5xx to 23,0xx.

I'm assuming you mean base clock set to 15xxMhz? What would the boost clock be set then something just little higher?

 

[turbogear]

I'm assuming you mean base clock set to 15xxMhz? What would the boost clock be set then something just little higher?

I did not set the frequency for P6 and P7 manually. Somehow setting manual value of frequency for P6 and P7 was not working properly in Wattman for me.
I left the Wattman on Frequency in % and increased the slider slowly in steps starting at 1% and doing benchmarking for each step.
I started at around 975mV with stock frequency settings and repeated until I got a stable and good value for voltage, frequency and power draw at the wall.

It was quite a time consuming task.

I am currently running 1075mV with Frequency OC of 2.5% and power limit set to 50%.
With 2.5% frequancy OC, the P7 state would be around 1671MHz. I don't know how much does the Wattman set for P6 in this case.
Voltage on P6 and P7 are both set to 1075mV.

With these settings, I am getting an average clock speed in stressful games/benchmarks of around 1620MHz with around 20W more power consumption over the Turbo mode.

Just to mention, I have custom liquid cooled Vega 64 card.

 

[londiste]

from my experience clock will go down when voltage is decreased. throughout my testing, i had the frequency set to 1632mhz (or whatever the default was there).
out-of-box, this resulted in 1440 mhz in games. firestrike - 22 468 @ 1488 mhz (graphics score).
+50% power limit - 1570 mhz in games.
voltage to 950 (in wattman) - 1550 mhz in games. firestrike - 23 439 @1570 mhz.

that clock is not really held. ashes of the singularity in particular stood out as a game that both drove clocks down (1300 in stock and 1420 when undervolted) and power consumption up (with 50% power limit definitely to 300+w for the card).

also, even in that last config, i needed fan speed at 3400 to keep 80+-2c temperatures.

memory does seem to easily run hotter than gpu when overclocked. memory itself can probably be overclocked to 1050-1090 mhz but when gpu load comes into play, these 990-1020 mhz frequencies is where it is at, at least with the reference/air cooling. watercooling should do miracles here due to lower temps (but not performance miracles really).

 

[kurosagi01]

So after 2 hours or so finding a setup im comfortable with at the moment here are my results so far for evening in attempt to run it at stock speed with lower voltage.
Stock setup on firestrike ultra are:
Total : 4927
graphics : 5320
physics: 7004
combined: 2466

 

[xkm1948]

So after 2 hours or so finding a setup im comfortable with at the moment here are my results so far for evening in attempt to run it at stock speed with lower voltage.
Stock setup on firestrike ultra are:
Total : 4927
graphics : 5320
physics: 7004
combined: 2466
View attachment 91838 View attachment 91839 View attachment 91840

What is your stability score? Since you have a full version of 3DMark. >98% in FireStrike Ultra and >99% in TimeSpy means you are good to go for long term stability.

 

[kurosagi01]

What is your stability score? Since you have a full version of 3DMark. >98% in FireStrike Ultra and >99% in TimeSpy means you are good to go for long term stability.

Didn't try the stress test ultra,ran one test at the same setup and got 74% which is not good lol.

Edit:
Interesting..i turned it to default "balance" and ran test again and got 94.8% which means according to 3dmark the balance setup isn't even stable either.

 

[xkm1948]

Didn't try the stress test ultra,ran one test at the same setup and got 74% which is not good lol.

Edit:
Interesting..i turned it to default "balance" and ran test again and got 94.8% which means according to 3dmark the balance setup isn't even stable either.

94.8% is way too low for long term use.


74%, well shit didn't see that. Welp good luck fine tuning!

 

[kurosagi01]

94.8% is way too low for long term use.


74%, well shit didn't see that. Welp good luck fine tuning!

Well i don't know whether to take it with grain of salt or not,i mean 94.8% at default settings is still confusing. Could be temperature making the difference? I bumped the fan speed upto 4000 max and i got 77% this time with the P6: 1070mv and P7:1100mv

Edit: Changed the P6 frequency to default 1536mhz and got 83.6% with fan speed set at max 4000 which is improvement but man is it loud..

 

[xkm1948]

Well i don't know whether to take it with grain of salt or not,i mean 94.8% at default settings is still confusing. Could be temperature making the difference? I bumped the fan speed upto 4000 max and i got 77% this time with the P6: 1070mv and P7:1100mv

FuryX bone stock, 99.8% FireStrike Ultra.

FuryX overclock 1065 core with default voltage . 95% stability.

FuryX overclock 1100 core +50mV , 97% stability

Just for reference

 

[kurosagi01]

Well i just managed 97.2% now with firestrike lol
P6: 1536mhz 1070mv
P7: 1632mhz 1090mv
fan speed set to max 4000rpm
power limit: 50%
Ideally would like run fans a lot quieter but i doubt thats possible with reference cooler?

 

[xkm1948]

FuryX does have AIO though. I would say your 97% is good enough for now

 

[terroralpha]

What numbers do people recommend testing first with undervolting and power limit at stock speed as i am new to the whole undervolting thing. Been reading different sites on user experience and all numbers vary since every card are different.

Source *from reddit amd group*:
https://docs.google.com/spreadsheet...beRiOTMsGeIkCDQUhXSBA/htmlview#gid=1964178281

for core, first try 1000mV across the board, and 1100mV for memory, leave core alone but OC memory to 980MHz for starters. increase power target to 50% (max). if your card can handle that, you're pretty much set. the memory OC should yield a 5% or better boost in FPS. power draw should drop significantly.

if all goes well, see how far you can take the memory in 10MHz increment, but note that gains from memory OC will dropped to near zero after 1000MHz for me.

 

[kurosagi01]

Update 8/09/2017 6:50am UK time
P6: 1050mv P7: 1050mv stock frequency
mem freq and mv default
Power limit: 50%
fan max speed: 4000rpm
Firestrike Ultra:

P6: 1050mv P7: 1050mv stock frequency
mem freq: 980mhz 1100mv
power limit: 50%
fan max speed: 4000rpm
Firestrike Ultra:


Pretty pleased with these results..going try reduce P6 and P7 hopefully to 1000mv like terroralpha recommended when i get home after work and gym and hopefully..play some damn games haha.

 

[kurosagi01]

Update 6:35pm UK time
Couldn't get the voltage down any lower without running into low "stability" score sadly..however when i set the voltage back to my earlier setup this morning i can't replicate that result at all.
So what could it mean? It can't handle it or i just got lucky from room temperature being on my side in morning as it was pretty cold and room temperature has increase as time passes.

 

[TheoneandonlyMrK]

for core, first try 1000mV across the board, and 1100mV for memory, leave core alone but OC memory to 980MHz for starters. increase power target to 50% (max). if your card can handle that, you're pretty much set. the memory OC should yield a 5% or better boost in FPS. power draw should drop significantly.

if all goes well, see how far you can take the memory in 10MHz increment, but note that gains from memory OC will dropped to near zero after 1000MHz for me.

Why 1100mv for memory its not noted(not just me many review and oc guides) to either do much or stabilise memory , with stck memory volts im good upto 1100mhz game stable and mire volts doesn't change that top limit for me atm.
My best oc varies 1729-1751 core 1100 memory GtaV etc stable ,i run it lower to save power though.

 

[kurosagi01]

Why 1100mv for memory its not noted(not just me many review and oc guides) to either do much or stabilise memory , with stck memory volts im good upto 1100mhz game stable and mire volts doesn't change that top limit for me atm.
My best oc varies 1729-1751 core 1100 memory GtaV etc stable ,i run it lower to save power though.

I just tried running the mem voltage at stock 1050mv at 980mhz and my system just crashed 5 seconds in firestrike lol.

EDIT:
So i think i'm done with undervolting now,wish i could reduce the fan speed to keep the stability but here we are..
P6: 1050mv P7: 1050mv stock frequency
Memory: 1100mv 980mhz
Fan speed max: 4200rpm
power limit: 50%
Firestrike Stability result: 97.7% (forgot save)

Here is my firestrike ultra benchmark score: