• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Vega 64 Overclocking & Undervolting information & guide

Soldato
Joined
30 Dec 2011
Posts
6,085
Location
Belfast
I have been extensively testing overclocking and undervolting my Vega 64 AC card. 17.9.1 Drivers, Windows 10, Ryzen 1700X at 3.8GHz, 32GB 3000MHz DDR4.

For clarity, my Vega 64 has been modded with a G10 Kraken (new holes drilled) and a Corsair H110 AIO. It maintains ~42c - 52c from lowest voltage tested to highest. There was no throttling on my AIO modded Vega 64 at any setting tested. At stock balanced mode using the reference cooler, core clocks would eventually throttle to ~1390.

Overclocking Vega 64 in the classic way, where you set a clock speed is not the easiest way to overclock. Vega core clocks are now more to do with the voltage you provide/allow it to use. All my testing was done by setting the P6 State to 1632 and P7 to 1682, then leaving them there and undervolting to see the results. Setting higher P7 state would result in instability. Also ensure P6 and P7 states are different, or you may encounter a bug where HBM speed drops significantly.

Here is the table showing various clock speeds attained at the various voltage settings.



As can be seen, going from stock balanced to the max OC tested, results in ~15% extra performance. I only tested one game but it was fine as a reference point. I could push further by using registry hacks, or flashing the LC BIOS but I won't :)

What is astonishing is the difference in power usage from lowest stable voltage to max allowed voltage. 153 W to 325 W, over twice the power used for 11% extra performance.

As noted in the table, setting states P6 and P7 voltages to anything lower than the memory voltage setting had zero effect. So if you are undervolting, make sure memory voltage is set to your desired minimum voltage. This is why I believe many reviews got their undervolting tests to show voltages no lower than stock. They were just doing it wrong and assuming Wattman was broken.

For reference, here is the setting I was using on the stock cooled card. With stock cooler, max sustainable speeds with no throttling and acceptable fan noise was 900mV, 1435/1020, 2300RPM, core temperature was 77c - 80c depending upon ambient. This would result in performance ~2.5% - 5% higher than stock because it eliminates throttling. 150 W compared to 220 W power consumption, a 30% reduction in power consumption and still faster than stock.
  • Note: On my Vega 64, highest stable HBM overclock on the reference cooler was 1020MHz.
  • The memory voltage label in Wattman is wrong, it does not undervolt HBM, it is more like a core "voltage floor", or power budget to be shared among core and HBM.
  • Your memory voltage tells the GPU what voltage it will NEVER GO BELOW. So this should always be equal or lower to your LOWEST voltage in P6 or P7.
  • As a consequence of the above point, it is useless setting your P6 or P7 to a voltage lower than memory voltage. It wont ever use it as memory voltage is the minimum.
  • At higher voltage settings your core clocks can fluctuate wildly without increasing PL to compensate.



My conclusions:
  • I found 1554/1100 at 1000mV to give a very nice ~10% performance boost over stock and with a ~10% reduction in power consumption.
  • Going any higher results in what are clearly diminishing returns for excessive power consumption.
  • These results are contrary to many reviews simply saying, "pushing the power limit to 50% at stock is a nice way to overclock". It is most certainly not and will result in your GPU seriously throttling very quickly.
  • On stock reference Vega 64, drop mV to 900 for all voltage settings and set PL to 50%. You will get a better performing GPU that does not throttle, has a very noticeable reduction in power consumption, noise and thermals. 30% reduction in power consumption makes Vega 64 a very efficient GPU.
  • Vega 64 is a decent overclocking GPU and by combining conventional overlcocking with undervolting, it is possible to increase performance while lowering power consumption.
  • Conventional overclocking methods alone do not work well.
  • Wattman is very useful once you work around the bugs.
  • The memory voltage label in Wattman is wrong, it does not undervolt HBM, it is more like a core "voltage floor", or power budget.
  • 950mV for all voltage states will allow ~1480-1500 Core to be maintained with 2400RPM fan setting.
  • 900mV for all voltage states will allow ~1420-1450 Core to be maintained with 2300RPM fan setting.
  • If you are undervolting and your core clock drops or fluctuates massively, you need to increase Power Limit. I use +50% to avoid any core clock fluctuations.
I hope this guide is helpful to any Vega 64 owners out there.
 
Last edited:
Very good detailed info on the bugs and general clocking although similar to Polaris it will help people.
Your Hbm voltage is not representative of the voltage the modules run at. If you are referring to voltage control in wattman then that doesn't undervolt the modules from 1.35v.

good work though

Thanks. I know the memory voltage label in Wattman is wrong, it doe snot undervolt HBM, it is more a "voltage floor".

I will add that as a note to my post.
 
I think what's probably happened is AMD have thrown volts at Vega to be more competitive with NVidia, if NVidia GPU's weren't so good they'd have probably released Vega at a more modest clock/voltage that isn't so power inefficient.

That was my first assumption but I now think it might be that and a combination of yields. My testing shows it is possible to massively undervolt (not sure what consumption at wall is) and still achieve better than stock performance, thermals and noise levels. I believe that not all Vega 64 cores will undervolt that much and would have failed AMDs "minimum standards". They may have set the bar for failure rate to low in order to improve yields?
 
I gave mine a run out on 900mv on the voltage settings, 50% PL and it was great. I've been trying that new WH40K Inquisitor game which is a very rough alpha, in places the card would run flat out in menus, or certain areas setting the fan into over drive. After the tweaks everything was fine. Nice and quiet too.

Yep, 30% reduction in power consumption makes Vega 64 a far more efficient GPU. ~30% less power consumption and ~5% faster than stock.
 
Power Limit essentially increases the power you allow the GPU to use by the % selected. At lower voltages it does not seem to have any noticeable effect, but no harm leaving it on.
 
So does that increase the voltage as well as the clock rate by a %? Or is a % increase of the overall power draw wattage? Sorry if these are daft questions :)

No voltage will be the same, it simply instructs the GPU to use an increased amount of power, which will allow the GPU to reach higher boosts. Power draw (wattage) increases substantially (from 220W to 310W at stock).
 
Do you have any idea why the first row achieves higher GPU clocks than Stock+50% power limit?
The only difference seems to be that memory is clocked higher on former.

Stock clock settings on P7 is 1632, while the manual settings I used were 1682. So a slight overclock on core clocks was used in row 1. Hence the slightly higher total wattage used.
 
Check GPU-Z to see what GPU only power reports? At 1000mV it should be around 205W. Is this using ROTTR, because less taxing games or benchmarks will possibly boost higher.
 
It's good to see people still finding this guide useful for their new Vega GPUs and thanks for the feedback from everyone. Like many have posted Vega is a very fun GPU to play around with as a Nerd :)

I have noticed the latest Adrenalin drivers do not seem to require the HBM voltage or "voltage floor" to be changed. I was wondering why my HBM speed was stuck at 500MHz and put the voltage back to stock 1150 for HBM and it fixed the low HBM speed, but also the GPU core was still only pulling ~160w - 180w game dependant.

I am using stock cooler and .900mV to keep fans at 38% which is around 2300rpm. This keeps the GPU at a solid 1440ish on the core and is actually a mild overclock of about compared to stock.
 
Try setting the P1, P7 states separately and leave HBM alone, start from P7 and work back until it drops as that will tell you what state the GPU is getting stuck on due to throttling. Monitor GPU-Z for power usage dropping after each change separately with a 3D load on the GPU. Also monitor HBM speed because I found that chaning the voltage for P1-P7 to manual caused HBM speed to drop to 800Mhz unless done one at a time.

If you get no power reduction, also set the HBM power to .950 and see if that makes any difference to both power usage or HBM speeds. I found dropping HBM to anything lower than .950 on latest drivers dropped HBM speeds to 500 or 800.
 
Back
Top Bottom