So how did AMD get Vega voltage so wrong?

Rifter

Lifer
Oct 9, 1999
11,522
751
126
So i finally upgraded my old RX480 8GB, to give it to my Girlfriend for her new build and scored a Vega RX 56 Gigabyte OC 8GB card for $500CDN on black friday.

I have herd that you can undervolt these Vega and Polaris GPU's instead of over volting when you OC, Unlike everything ive ever known, seems so backwards to me. And since i got a launch 480, one of the first to get to canada, it undervolted very very poorly and overclocked very poorly(its a MSI OC card from factory, was pretty much maxed out in its stock form) so i thought all this under volting hype was just that, hype, and you needed a golden chip to get any real effect. Turns out i may have been wrong and just lost the silicon lottery with my 480.

So I install the Vega 56, leave it at stock settings to get a baseline(1500Mhz, 1.2V) and fire up the benchmarks and run unigine superposition benchmark for two hours to make sure its fully at temp and take a look at whats going on. Its pinned at 100% GPU usage so thats good. Its drawing around 460w total system power(this is close to 100w higher than my RX 480), cards running at 75c GPU temps with VRMs and Memory being a few degrees hotter, pretty much inline with the reviews of this card. It maintained average clocks of 1400-1425Mhz, disappointing really with a 1500Mhz max clock that it only maintained for first few mins until it heated up. It is quiet even with the fan at 3800-4000rpms, GPU is by far the most noise in the system but its not that bad, people complaining about vega being loud must be referring to the reference blowers.

So then i try my hand at under volting and mild OC, not expecting much at all.

Well after a few hours of working my way down the voltage scale i find my card will run with only 1V, totally stable, as well as boost the Turbo core clock to 1600mhz. This is a rather large reduction from its 1.2V stock voltage, and with a 100mhz clock increase to boot. I have no idea whay AMD would run these so over volted from factory, its insane, AMD dropped ball really bad on Vega voltage, making them look way worse in reviews than they actually are when you use sane voltage levels.

With its new voltage After 2 hours of unigine superposition its drawing only 395w total system power, a 65W reduction. Its maintaining 1550Mhz clocks, its running at 65-67c, 8-10C reduction in temps, and is much quieter with the fan running about 500-600rpms less than at stock voltage.

Im going to see how far i can push the clocks now, but honestly just lowering the voltage and a 100mhz bump has already given me great gains, much much more than i could ever accomplish with my RX480.

Must admit im pretty surprised AMD would push these so hard on stock voltage for no good reason, even if some chips do need the voltage, you would think they would build in some Voltage logic circuit or something to under volt the chips that can do it..
 

n0x1ous

Platinum Member
Sep 9, 2010
2,572
248
106
Yeah it would be nice if they could tune them in some automated ways or something. I have 4 vega64 cards. (3 ref liquid, 1 ref blower) the blower and 1 of the liquid both came at 1.2 default and I can shave off 100mv on both and get a nice improvement, but one of the liquid has 1.25v default volts and wont even run at stock clocks with anything less than that!
 
  • Like
Reactions: UsandThem

Rifter

Lifer
Oct 9, 1999
11,522
751
126
Yeah it would be nice if they could tune them in some automated ways or something. I have 4 vega64 cards. (3 ref liquid, 1 ref blower) the blower and 1 of the liquid both came at 1.2 default and I can shave off 100mv on both and get a nice improvement, but one of the liquid has 1.25v default volts and wont even run at stock clocks with anything less than that!

Im sure alot of it is silicon lottery, but with so many being able to be undervolted im kinda surprised AMD hasnt released a Vega refresh with binned chips and much lower voltage.
 
  • Like
Reactions: esquared

psolord

Golden Member
Sep 16, 2009
1,875
1,184
136
So i finally upgraded my old RX480 8GB, to give it to my Girlfriend for her new build and scored

It was a nice story up to that point! xD

Joking aside, remember charlie bashing nvidia for going high current low voltage a few years back? Maybe it was not such a bad idea on nvidia's part after all!
 
  • Like
Reactions: UsandThem

cfenton

Senior member
Jul 27, 2015
277
99
101
I'm not sure what else they could do. They need every GPU they sell to work out of the box. If even 1 in 1000 won't work with less than 1.2v, then they need all of them to run at 1.2v. My understanding is that Vega has been difficult to produce for AMD, so they are probably forced to use every GPU that is even passable.
 

Rifter

Lifer
Oct 9, 1999
11,522
751
126
I'm not sure what else they could do. They need every GPU they sell to work out of the box. If even 1 in 1000 won't work with less than 1.2v, then they need all of them to run at 1.2v. My understanding is that Vega has been difficult to produce for AMD, so they are probably forced to use every GPU that is even passable.

I get that, but CPU's have had great voltage controllers for years which can adjust voltage on the fly based on different parameters, why could this not be built into vega GPU so at least the vast majority of chips could be run at a reasonable voltage. I mean this is a night and day difference, the vega launch reviews make it look like a horrible hot power sucking mess, but the undervolting reviews make it look much more competitive. AMD could have saved alot of bad press by either binning these chips alot better, or building in some voltage control.
 

Hitman928

Diamond Member
Apr 15, 2012
5,180
7,631
136
Have you checked to make sure the voltage is actually being applied? If I remember correctly, when I was undervolting Vega, it seemed to stop actually applying lower voltages at some point even though I was setting them lower and lower. It may have just been a bug early on as I haven't played with it in a while but I would double check the voltage at load just to make sure.

As far as AMD setting them too high, this is true. My guess is the reason it has been this way since Hawaii is lack of $$$. Yes it's true they want to use as many dies as possible, but fine tuning voltages takes a lot of validation time. That's a lot of man hours which means a lot of employees working on it to get it done in a timely matter. I'm guessing with all the cut backs AMD had to do, they just didn't have the manpower to fine tune Vega. Ryzen is actually the same way, I've seen some people running stock speeds on even the highest skus with significant undervolts. It's just more pronounced when the IC is using ~220 W and can't even maintain highest clocks versus full clocks and ~105 W.
 
Last edited:

Rifter

Lifer
Oct 9, 1999
11,522
751
126
Have you checked to make sure the voltage is actually being applied? If I remember correctly, when I was undervolting Vega, it seemed to stop actually applying lower voltages at some point even though I was setting them lower and lower. It may have just been a bug early on as I haven't played with it in a while but I would double check the voltage at load just to make sure.

GPUz and wattman both seem to think its working and the fact that my power usage is dropping by almost 70w would back that up.

There is a point where it stops dropping when you input lower core voltage values, this point is the HBM memory voltage, the GPU core voltage can not be set to lower than the memory voltage or it will not apply the changes.
 
  • Like
Reactions: ElFenix

ElFenix

Elite Member
Super Moderator
Mar 20, 2000
102,414
8,356
126
Have you checked to make sure the voltage is actually being applied? If I remember correctly, when I was undervolting Vega, it seemed to stop actually applying lower voltages at some point even though I was setting them lower and lower. It may have just been a bug early on as I haven't played with it in a while but I would double check the voltage at load just to make sure.

65 watt reduction in power consumption seems to back up that lower voltage is getting applied.
 

Hitman928

Diamond Member
Apr 15, 2012
5,180
7,631
136
I'm sure lower voltages are being applied, just asking if he checked that it's really down to 1V. From what I remember, mine stopped applying voltages lower than like 1050 mV or something. But again, this was on the earliest drivers which were a little buggy when it came to Wattman.
 
  • Like
Reactions: happy medium

Phynaz

Lifer
Mar 13, 2006
10,140
819
126
They didn't get it wrong - I'm sure it's 99% correct. For AMD's requirements.

Don't assume AMD's requirements and yours are the same.
 

Rifter

Lifer
Oct 9, 1999
11,522
751
126
I'm sure lower voltages are being applied, just asking if he checked that it's really down to 1V. From what I remember, mine stopped applying voltages lower than like 1050 mV or something. But again, this was on the earliest drivers which were a little buggy when it came to Wattman.

Its still an issue, and its 950mv, or at least my memory on my card was set to 950mv stock, and you cant set core voltage lower than memory voltage.
 

Hitman928

Diamond Member
Apr 15, 2012
5,180
7,631
136
Its still an issue, and its 950mv, or at least my memory on my card was set to 950mv stock, and you cant set core voltage lower than memory voltage.

Yes, that was also true on Polaris, but what I'm talking about was a higher limit than what the memory voltage was set at. Like I said though, Wattman was super buggy with Vega when it first came out, the issue was probably fixed by now, I'll have to give it a try and see if I can set it lower now.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
Maybe this is one of the areas that the CPU guys transferred to the GPU division will work on. The power management of their GPUs are way behind the new CPUs and also Nvidia.
 

Thunder 57

Platinum Member
Aug 19, 2007
2,647
3,706
136
There will be a feature in the AMD drivers soon that will help with this. Sounds rather useful. I know we can undervolt right now, but maybe there will be a way to automate it or make it easier?

"What’s more, the company is actually one-upping NV over here and including undervolting support as well. Some Radeon cards like Vega have known to be overvolted for stability (and ease of QA) purposes and many chips can actually run at the stated power levels just as happily on lower voltages. This undervolting support will allow users to minimize the power draw from their Radeon card (especially useful if you are running a mining operation – not that they are very profitable right now). "

https://wccftech.com/amd-hosting-pr...adeon-drivers-update-adrenaline-2019-edition/

Take it with a grain of salt considering the source, but it sounds very believable. I would be very interested in such a feature.
 

Qwertilot

Golden Member
Nov 28, 2013
1,604
257
126
That's a real fig leaf vs actually setting the factory voltages to something sensible though.
 

happy medium

Lifer
Jun 8, 2003
14,387
480
126
They didn't get the voltages so wrong, they are letting the consumers do there own quality control. Most
Cpu's, gpu's made today can be undervolted, its nothing new.
I don't see why this is such a big deal.
 
Mar 11, 2004
23,031
5,495
146
Have you checked to make sure the voltage is actually being applied? If I remember correctly, when I was undervolting Vega, it seemed to stop actually applying lower voltages at some point even though I was setting them lower and lower. It may have just been a bug early on as I haven't played with it in a while but I would double check the voltage at load just to make sure.

As far as AMD setting them too high, this is true. My guess is the reason it has been this way since Hawaii is lack of $$$. Yes it's true they want to use as many dies as possible, but fine tuning voltages takes a lot of validation time. That's a lot of man hours which means a lot of employees working on it to get it done in a timely matter. I'm guessing with all the cut backs AMD had to do, they just didn't have the manpower to fine tune Vega. Ryzen is actually the same way, I've seen some people running stock speeds on even the highest skus with significant undervolts. It's just more pronounced when the IC is using ~220 W and can't even maintain highest clocks versus full clocks and ~105 W.

The thing is, you'd think they could spare it some when it was making them look worse than they were (leading to losing market share) or they'd tell the AIBs so they could tweak it themselves to show big gains over the reference cards. They clearly recognized potential to use that for segmentation, as seen with the Fury Nano.

I think there were multiple issues. One was that their GPU designs were built for certain products for certain customers, and then they had to try and push clocks to compete with Nvidia in dGPU, and they didn't have the engineers to be able to work on the chips to make them more suitable to higher clocks they'd be pushing in consumer dGPU stuff, and then they didn't have the engineers to bin them much after they got produced (it seems fairly automated or pretty loose, assuming to maximize number of chips). And I think the people running the GPU division were focused on other issues (seems like Raja was focused on the software, which is another area that was limiting AMD's cards potential, as DX12/Vulkan showed, there was a lot of potential there going untapped; and Vega seems like it might be more limited because they didn't even get some of its best features working).

I don't think I've seen anything too crazy and I've seen most people need to up volts to get just an extra 100-200MHz boost on Ryzen. Quite a few people have even remarked that they got the best performance by just letting the chip manage clocks itself. I've seen very few people that can't significantly improve things on AMDs GPUs, with a 50-100mV reduction.

Yes, that was also true on Polaris, but what I'm talking about was a higher limit than what the memory voltage was set at. Like I said though, Wattman was super buggy with Vega when it first came out, the issue was probably fixed by now, I'll have to give it a try and see if I can set it lower now.

Its been a little while, but if I remember right it was I think it was 1.0V or something, and it was due to the voltage for the HBM setting the minimum voltage you could hit. Wattman would let you set it lower (think down to like 850 or maybe even 800) but below the one threshold it wouldn't actually improve things (but you'd only know it if you actually measured it in other ways). You could drop further, but you had to tweak things with the HBM voltage (which I think took BIOS modding or some software hack) I think some got down to like 825mV. It was low enough that they encountered another problem, and that was that when idle or at low manually set clocks, the voltage would cause some instability or something. I think I noticed this on Polaris as well. I had random restarts and then later it'd just crash the GPU (which Windows could soft reboot so it wouldn't even crash Windows) when the GPU would have low utilization - like single digit % when playing back videos - and I think it was due to having too low of a base voltage combined with low base clocks in the earlier power states. After I tweaked that a bit it stopped happening. I could always tell though as the crash would reset the clock/voltages and it'd make the room noticeably warmer and the GPU would be making more noise - even at idle.

I don't believe it has been fixed in Wattman for whatever reason. Supposedly you can see pretty good performance increase by dropping the GPU volts, setting it to like 1600 max (with like the top two states being high like 1500+) and then finding the sweet spot for the HBM volts where you can push clocks higher on it. Sounded like a lot of fussing though.

I do believe they fixed the issue where you had to set Power Limit to max (50% I think) while undervolting otherwise it wouldn't do much or was causing an issue or something. So now you can just undervolt (I do that and then save the profile, and then tune from there). I don't even know how much I ended up undervolting but I tweaked the clocks as well (think I have 1200MHz at 1025mV at P6, 1100 at 950mV for P5, and then I actually downclocked just a bit for P7, to 1275 at 1075mV - which I believe stock was 1290 at 1150mV?).

Maybe this is one of the areas that the CPU guys transferred to the GPU division will work on. The power management of their GPUs are way behind the new CPUs and also Nvidia.

Hopefully as its so bad it actually could impact them beyond consumder dGPU. Its so bad that Microsoft took to binning/optimizing themselves. If they hope to get their GPUs back on track, they absolutely need to be improving these aspects.

They didn't get the voltages so wrong, they are letting the consumers do there own quality control. Most
Cpu's, gpu's made today can be undervolted, its nothing new.
I don't see why this is such a big deal.

Yes they did, because the less voltage they ran, the less power they'd use, the less waste heat they'd generate. Heck, it probably would've prevented the power issue that happened with Polaris at launch.
Very few are that far out of range or see that much benefit from it. And few that I've seen can gain performance while undervolting (you usually tradeoff absolute performance to undervolt, or you overvolt to gain performance at the expense of efficiency; with Intel and Nvidia it seemed like you could just overclock at stock voltage and not take too big of a hit in efficiency while getting increased performance, they were less fussy - which because they seemed to be better optimized from the get go, you didn't tend to gain a lot from undervolting either, and I think it made their overclocks look better).
 

happy medium

Lifer
Jun 8, 2003
14,387
480
126
DarkSwordsman17,
So you think AMD got it wrong for the past 4 or 5 years?
People were undervolting Fury, then Polaris, now Vega.
They do it because not all AMD chips are created equal enough, so they use the highest common working voltage, just like any other chip.
Like I said this is nothing new.
When you have chips that use 100 watts more than the competition it becomes a big deal.
Example, lets say the rx580 with equal performance to a gtx1060 used 20 watts more, who would be looking to undervolt a 580?
Know one would.
 

Elfear

Diamond Member
May 30, 2004
7,096
640
126
They didn't get the voltages so wrong, they are letting the consumers do there own quality control. Most
Cpu's, gpu's made today can be undervolted, its nothing new.
I don't see why this is such a big deal.

Most CPUs and GPUs can be undervolted but you don't see the huge efficiency gains like you do with AMD GPUs the last few years. I believe, as someone else stated, that this is due to AMD's lackluster budget which inevitably caused their validation process to be underfunded. This resulted in their good, average, and poor chips being run at the same voltage in order to ensure stability.

DarkSwordsman17,
So you think AMD got it wrong for the past 4 or 5 years?
People were undervolting Fury, then Polaris, now Vega.
They do it because not all AMD chips are created equal enough, so they use the highest common working voltage, just like any other chip.
Like I said this is nothing new.
When you have chips that use 100 watts more than the competition it becomes a big deal.
Example, lets say the rx580 with equal performance to a gtx1060 used 20 watts more, who would be looking to undervolt a 580?
[No] one would.

The tweakers would still undervolt but point taken, undervolting became more popular when AMD's chips were consuming more power than their Nvidia counterparts.
 
  • Like
Reactions: happy medium

Headfoot

Diamond Member
Feb 28, 2008
4,444
641
126
Eh, undervolting got popular in a significant way when undervolting actually resulted in the cards running at higher average clockspeed due to reduced load on the cooler/fans. Like any other tweak, it only gets really popular when it makes it go faster.
 
  • Like
Reactions: burninatortech4