RX Vega Owners Thread, Tests, Mods, BIOS & Tweaks !

You can’t flash a stock card BIOS on a custom PCB card, since the manufacturer can change the BIOS’s parameters.
You can tweak it as any other Vega with Wattman/OverDriveNTool/PowerPlay tables.

 

I’ve been using it but looks like I’m voltage limited at 1.15V. Any way to increase this with PP?

 

Hmm, i can set it up to 1.250mV
Now im on 1.137mV for 1757MHz

 

I can set 1200mv in wattman but according to GPUz it will never actually hit that. Highest it will ever go is 1150 regardless of 242% power enabled by PP. I get hard crashing when pushing higher than 1650 core and 1150 memory but is stable when I do either or so for sure running into voltage or power limits.

 

That's about right, vdroop should cause around 30-50mv droop at 1200mv set value under load. It should hit 1200mv at idle though, if that's not the case the setting is not taking properly.

I noticed that sometimes my voltage setting won't change until I change a MHz setting for core or HBM. Then it sets the correct core mv setting I have. Not sure why that occurs but it's a frequent occurrence for me. I always double check my voltage settings take.

 

Yeah, the problem is that by default p5 is 1474MHz/1100mV on my card and without a significantly increased power limit it will never go past that, meaning my p6/p7 settings are irrelevant.

Since I couldn't leave well enough alone I've spent the last few days exploring SoftPowerPlay modding, and doing a few hundred Time Spy runs.

Conclusions first I suppose...

• My seemingly stable 950MHz HBM2 overclock wasn't, a number of stress tests later and I'm at 875MHz.
• Using the RX Vega Sapphire 56 Pulse UV reg softmod OnnA provided in the OP and the default settings of my card as a base I've come up with a 165W table that is Time Spy and Furmark stable and gets me to ~6930 graphics score in Time Spy. Reaching a high of ~1550-1560MHz in GT1 (@975-994mV) and ~1605-1615MHz in GT2 (@988-1001mV).
• Using the same softmod with +9% PL (180W effective) gets me to ~7245 graphics score in Time Spy with a high of ~1630MHz in GT1 and ~1645MHz in GT2 @1000mV or so.
• With +9% PL I'm not actually 100% power limited in GT2, hitting ~167-170W rather than 180 for part of the test, which means I need more voltage to get higher clocks. However, if I increase voltages for p6 and p7 I'm actually reducing performance under the power limited parts of both tests which led me to conclude my sweet spot is probably something like 170-175W and I've settled on the 165W softmod +5% PL.

Honestly playing around with this stuff has been great fun, I haven't actually gamed more than ~30 minutes since I got the card but I already feel I've gotten my money's worth!

Tips for anyone else wanting to play with softmods, bearing in mind that I'm by no means an expert.

• Every single one of my instability issues has been due to memory. Under any reasonably power constrained scenario the card seems to only clock as high as the given voltage will allow, as such I haven't suffered from any core instability at any point.
• If you want the card to reach idle clocks while, uh, idle then the p1 power state must be higher than the p0 one. If you set the power level for both to the same number then memory clocks will idle at the highest strap that uses p0 as a baseline. For example, in the softmod provided in the OP p0 and p1 are both set to 800mV, making the memory idle at 700MHz as that state is tied to p1.
• I strongly suggest leaving the p0/p1/p2 power states at default voltages, 800/900/950 mV in my case, as these seem to tie in to power island voltages for the SoC as a whole and messing with them seem to have negative consequences on performance. Not 100% on this but observations seem to indicate that while lowering p1 (tried at 805/825/850 mV) does lower power consumption at every performance level it also disproportionally drops performance, I'm assuming from the SoC being power starved but that's just speculation. The top memory state is tied to p2 and messing with that may, or may not, have impacted my memory stability.
• Similarly I'd keep the memory states tied to their default voltage states (p0 for 167 and 500 MHz, p1 for 700 MHz and the 800 MHz top state to p2). I can't be sure that some of my instability came from tying the top state to p4 as opposed to p2 but since there's been no performance advantage to play with the power of those early stages anyway it's best not to touch them.
• According to what I've read online you can't change the HBM2 memory voltage (default 1250mV on the 56) with softmods, the number is ignored and I haven't mucked about with it since all my instability was due to memory anyway.
• I've used VEGA SoftPower Table Editor (requires GTK# for .NET to work) to edit the softmod file.
• The idle core voltage seems tied to p1, based on it hitting 850/900mV depending on what I set p1 to. Occasionally this drops to 775mV, usually as a result of an unstable overclock resulting in a reboot (my guess it's a "safe" baseline, similar to how WattMan forgets all your settings after a crash) but I've also seen it happen when coming out of Sleep mode and when messing with WattMan profiles. When it drops like that your performance will be impaired and I recommend rebooting again.

 

You need to remember that Now (18.11.2) Adrenaline has new A.I. controlled Voltages & Clocks.
Im observing now a little lower tW (Thermal Watt) on Average thanks to this (Refined) feature.

If you don't hit 1.150v then it's not needed
Always try to find Lowest V possible -> or run Balanced in WattMan Tab.

 

Vega 56 faster than GTX1080. What madness is this?

 

@Exodite You can disable P0-P6 in Wattman or OverdriveNTool. This in turn allows use of much lower voltage for core/HBM (works for both) when doing an undervolt. As you can see from the image below it's very possible to run less than 1100mv



For reference, I think there's still legs in your card yet.

Currently my top score in Timespy is 8446 with my Vega64. That was done with 1752MHz core and 1135MHz HBM set. I should probably do a re-run with 1752MHz Core and 1180MHz HBM set as I know my card can bench 1180MHz.

Another thing to try would be a 64 BIOS on the 56 and see if the voltage can be used for the HBM. 1.35v vs 1.25v may provide more stability, though equally it could just cause instant crashing (if so switch BIOS on the card to the low power one and load back in, then re-install the 56 BIOS to the high perf position).

 

Best for Vega ($439)

-> https://www.samsung.com/us/computin...ming-monitor-with-quantum-dot-lc27hg70qqnxza/

 

Cheers, thanks for the new info!

That made me look up the documentation for OverdriveNTool, as well as get a newer version.

It kinda confirms a feeling I've got while playing around with the p-states too, that the card extrapolates clock and voltage steppings based on the deltas between the same for the various p-states. For example, I currently have p5 @ 1450MHz/975mV. If I increase it to 1475MHz/975mV, something I know the card can easily handle, it ends up reducing my max clocks while effectively running within p6 level. I might get 1610MHz rather than 1630MHz, say. Finding the perfect balance of clocks and voltages for the various p-states could probably optimize performance further but it's likely to be incredibly time consuming as it all feels very interdependent.

I have no doubt I could push the card further, at least when it comes to core clocks, but my intent has pretty much been to optimize performance at the power level I'm willing to give it rather than get max performance. I'm using this as a 24/7 profile and since I'm spending 90% of my time not loading the GPU I'm not willing to forsake the lower p-states. Hence why my first changes to the softmod BIOS in the OP was to get idle clocks to work.

As I'm doing ~1640-1645 MHz at ~1000mV I don't doubt I could do better, especially if disabling the lower states reduce power draw further, and I might look into that for benchmarking purposes if nothing else. But not right now, I'm pretty much at maximum saturation for Time Spy runs that my body can handle.

I don't know how reliable GPU-Z is with its information but it claims I've got Samsung HBM, a rarity for the 56 if I understand things correctly, and if that's true I imagine my memory could do better with higher voltages. Though I'm loathe to try that I admit, I know it's hard to mess up too badly with a dual BIOS system but I'm not willing to risk it regardless as it's a bit outside my comfort zone - especially for a performance gain I'm unlikely to use outside of testing or benchmarking purposes.

Anyway, thanks again!

 

During idle my voltages are much lower due to downlocking so I actually never see higher than 1150mv from GPUz. Is there a way to set LLC like for cpus to avoid vdrop? I can undervolt without changing MHz setting and it’ll apply no issues.

 

If you do have Samsung memory, there's no harm testing a 64 BIOS. You'll still have the lower stream processor count, but the HBM voltage boost could allow hitting 1100-1200MHz and in turn open up a huge amount of memory bandwidth to unblock that core.

You flash it in Windows using ATIflash 277 version and if there's an issue you just flick the physical switch. Load back in, then while in Windows flick it back to primary and re-flash.

Just make a copy of the original BIOS first before flashing over it

I've tried about 6 different ones so far and I settled on the 8774 LC BIOS. It's not one that works for all though and one of the Air 64 BIOS maybe a better fit for your card.

 

vdroop isn't a bad thing, it prevents damage to the GPU for a start! No need for LLC just set a value 50mv higher than your targeted value. The setting in Wattman/OverdriveNTool is still intelligently managed by the card anyway, it's not a hard setting more a target like the Power Target or the Temp Target.

Something people don't seem to know when they get hold of these cards is that you can disable lower P States. That in turn prevents downclocking when you need full bore. If I allow it to run stock it can cause stutter in some games that aren't GPU heavy as it swtiches between high and low P States, so for all games I run one of two profiles. Low power for games that aren't GPU dependent, using about 70-120W of power and then mid power which is set to 1100mv on core and 1050mv on HBM. I always run 1750MHz on core and 1100MHz on HBM and a 100% Power Target (using a softPPTable registry edit) to remove bottlenecks and use the voltage target to manage power, clocks and temps with a single setting.

 

Thanks for the reply! Forgive my ignorance but what is tW?

Also I’m beginning to think that my instability is caused from hotspot temps exceeding 95C even when the rest of my temps are blow 55C. Is there any word on how this hotspot temp is calculated?

 

Thanks for the reply! What are your temps when clicked that high? I can’t seem to get anything over 1700 stable and I think it’s tied to hotspot temps which I can’t seem to get undercontrol even with a custom loop.

As for modding the PPT did you only change the power percentages or did you also muck around with the current?

 

tW (Thermal Watt)
Power & Heat

 

Tbh gotta play around with my vega settings. I can see there are some things to get the performance to around 24000-25000 in 3dmark firestrike. I am at 23800 now with 1550 on core and 1100 on hbm2 with some 1050mv and 1075mv on p6 and 7.

 

Temps are pretty good for me as I get a 6C over ambient idle temp and top out around 55C with a 35C ambient. If I do benching with max performance settings (1750MHz/1250mv, HBM 1180Mhz/1100mv) I can hit 62C with my standard fan profile, though with max fan speeds it reduces that a fair bit. However setting the Clock speeds high doesn't 100% equate to getting those speeds as the variable controlling the actual core speeds is the voltage. At 950mv I'm getting between 1440MHz and 1530Mhz actual depending on load, very low power draw too with a max of around 140W and in iRacing I get around 70W usage as GPU usage is around 40%. Somthing like Ark which has 100% GPU but less load than a benchmark hits around 120W-130W. All wattage values are from GPU-z so not 100% accurate, but correlate to each other well.

Don't worry about the Hotspot temp. It's an erroneous temp which doesn't really affect anything. Focus on core and HBM temps and make sure they are stable and react as expected.

I also should mention I switched from paste to liquid metal as the pastes just don't react fast enough to the temp changes. Just make sure to use something to cover the components around the core/HBM modules and you'll be good.

Vega64 can hit up to 28000+ so definitely some headroom there mate

 

Yup, first try to hit 26k+ then 27k+
As for 28k -> We need Fast Driver & and good OC.

Note:
Try to find Best UV for 1600MHz

Here some Tips: