Return to Level1Techs.com

It's time to get (R9 FURY)ous

amd
amdgpu

#1

In this thread, I plan to chronicle the trials and tribulations of overlocking my Sapphire Nitro R9 Fury OC+.

While it is the higher binned variant of Sapphire’s offerings, clocking in at 1050mhz, it already near the maximum limitations that its silicon has to offer - not that the limitations imposed by the architecture, and the functions of Powertune help. But, I am not an electrical engineer that can pontificate and explain how certain design choices can limit the performance of this card. I’m just a guy that wants to tinker and get more power - correction, MOAR POWR!!!1111oneoneone

This will be a multi-post, and probably multi-stage, process, and I hope others join in with their findings, or suggestions, as I make this journey.

Prelude and Setting the Stage

Up until this point, I have never bothered overclocking my card. Why? Overclocking the Fury line-up of cards is, by and large, a fool’s errand. While performance may scale with an increase in clock speed, power consumption required for increased speeds can be described as absurd; for an 5% increase in clock speed, with comparable increase in gaming performance, your power consumption may increase by as much as 25%, depending on amount of voltage you have to feed the card.

Far more success and satisfaction can be had by attempting undervolting. As laid out in this Tom’s Hardware article, there is a lot of efficiency to be gained by undervolting. Depending on your luck, for minimal performance hits, you may end up shaving almost 100 watts of power consumption, which is impressive to say the least.

But we’re not chasing efficiency. This is not a rational venture. This is exploring the limits of what my card has to offer.


Test Setup:
R7 1800X
Gigabyte X370 Aorus Gaming K7
16gb of Corsair 3200mhz DDR4
EVGA 1050 SuperNOVA GS


Preliminary testing has not been favorable in the least.

I figured using Sapphire’s overclocking solution, Trixx, would be an fun route to take, and went to their website to download the latest version, v6.4.0, which was released 5/8/17. This turned out to be a frustrating assumption, as simply pressing the “apply” button on the interface causes the card to be locked in its lowest power stage. It even overrides powertune’s capacity to affect the clock rates, and attempts to “free” or modify the clock settings can not by done with other overclocking programs once this “bug” has been effected. Resetting defaults via the program’s interface has no effect; the same with closing or restarting the program. Once effected, the only way to undo the bug is to do a system reboot.

Still striving to try something different, I explored AMD’s overclocking solution: Wattman. Having access to the various power stages, and being able to manually input the voltage and clock rates for each individual stage was an interesting and fairly straightforward feature. The basics are there, and fairly intuitive to use.

However, I came across a feature most inconvenient. While I could increase the power limits by up to +50%, and could manually input the voltages, I could not raise my voltages beyond 1.250V - which is the stock voltage for my card’s final power stage when it operates at 1050mhz.
This would be acceptable if it were not the fact that, through trial and error, I have discovered that my card, with the power limit set to +50% and the fans screaming at the full 100% fan speed, can not maintain 1100mhz.

To achieve a mere 5% increase in clock speed, I will have to increase my voltage.

Wattman, while a simple and convenient tool to use, lacks the feature I need most. I give it props for actually working, but for this job, I had to download and install the old standby: MSI Afterburner, v4.5.0, released last month.

Goals and Looking Forward.

My starting goal will be to see what it takes to achieve and maintain a 5% increase in clock rate.

Afterwards, I will explore the limitations of the (air) cooling solution that comes with the card. I don’t have the money for custom water cooling solution (I don’t think one even exists for the Sapphire Nitro series of the Fury) and am not quite crazy enough to venture into the world of LN2. However, I am told, both by those with experience and by what I have read, that safe voltages are a non-issue with Fury - even those who resort to water cooling are thermally limited before they hit unsafe voltages.

Once core clock limitations have been explored, I will see what can be achieved with the HBM, then perhaps explore undervolting.


#2

well with fury, i’d say limit is quality of silicon… its bad; so to get decent oc you’d need too much voltage.

1125MHz 545.45MHz is obtainable on my fury-x, but i keep it on stock anyway since gains are too small to matter, and gains in temp are too big (especially on vrm). Best way is to edit bios to apply the clocks as highest stepping. Often you may find that lowering voltage may bring you up, as vrm heats up fast.

good luck out there.


#3

They oc a little but mostly a waste of time. Critical to keep the temps down. Somewhere about 70c things start getting weird. I dial the temps to about 55 ish underload and call it a day.


#4

My short term goal has been achieved!

The Proof is in the Pudding

I was actually surprised at the relative ease I was able to make this goal, but, perhaps, the hoops I had to leap through to reach this point has been so arduous that my expectations of difficulty were too high.

I often forget that overclocking is generally pretty simple, and the hardest part is enduring the time spent running stress tests. But, I am getting too far a head of myself. Let’s talk about methods and findings.

Methods and Findings

My testing methodology is far from anything complex or comprehensive. To offer some backstory, I used to almost exclusively used Furmark. However, with the release of the Hawaii-series (R9 290X, etcetera) and, afterwards, Tonga and Fiji, AMD has been making additional tweaks and additions to Powertune, which already made messing with the clock-rates sometimes a curious oddity. Furmark hates Powertune, and Powertune hates Furmark, and together they will gladly turn your VRM circuitry into a No Man’s Land of tears and sorrow. Remember when Powercolor came out with its Red Devil RX 480, and warned against using Furmark? It wasn’t because Powercolor necessarily created a product with a sub-par VRM, it was because Furmark will push your circuitry to limits beyond where even the most demanding games will - and not in the good way that stress-testing software, by its very nature, should.

Anyway, for time being, I have fallen on 3dMark’s Firestrike stress test utility. Unigine’s Heaven, and Superposition, stress test utilities are other ones I dabble in, but then I don’t get to share nifty results with the rest of you fine people.

For the actual method, I worked things a bit… backwards. Traditionally, you start with a given voltage, then work your way up with your clock rates until you reach instability with your testing. Once you reach instability, you add some more voltage until you are stable again, then rinse repeat.
In this case, I did the reverse: I set my clock rate to 1100, and increased my voltage by 0.006 V steps, running a stress test with each step. (Why 0.006? Turns out, those are smallest steps that the Controller IC, which tells the VRMs what to do, can handle.) The proof of the stability, linked above, was reached when my voltage offset was 0.030; theoretically giving me a voltage of 1.280. (Don’t worry, thanks to Powertune, actual average voltage was different.)

This is a good start, as my total voltage offset available to me with MSI Afterburner is 0.100V, which means I have room for a potentially higher clock rate. I can already hear the VRM screaming for mercy.


Settings:
Core Voltage: +30mv
Power Limit: +0
Core Clock: 1100
Memory Clock: +0
Fan Speed: 62% ; 2500 RPM



#5

Diminishing returns is a definite reality, and my VRM does not care for it one bit.

In the name of trying to safe time, I got brave, and daring. First, I set my fan speed to 100%, sending them screaming at approximately 3600 RPM. It sounds like a mini-jet engine inside my case, which is in stark contrast to the generally quite operation that the GPU handles. Secondly, I doubled my voltage offset, to +0.060, and set my clock rate to 1150.

I like to live every week like it’s shark week.

Stress testing did not go terrible. I never had a hard lock-up, although there were some program crashes with 3dMark. Twice, I had to reboot the computer because the P-State was stuck at maximum after the program crash. Manipulating via MSI Afterburner had no affect, so a system reboot was required to reset my voltages and P-states. Studying the data, I learned that +0.060 V causes is too much heat for the cooling solution handle - Shortly before the crash, the GPU temperature begins to ebb off, which can be interpreting as the VRM throttling itself due to overheating. The crash comes with a sharp spike up in voltage, which then Powertune forces down to stock settings. (This is a feature, not a bug.) However, this is just an interpretation - I’m going to have to find a program that taps into VRM thermal probes (I think AIDA64 does?) to confirm.

Either way, I decided to attempt creeping up the voltage from +0.030 V to see what might be the limit for my VRM. Trying to run 1150 with that voltage led to pretty hard crashes, leaving me to believe that 1150 certainly requires more voltage. +0.042 V offset almost completed a full run for a stress test, but crashed with the “VRM Throttle” symptoms.

Core Temperatures hovered at around 65 C.

More tinkering is required.


#6

Last I checked I couldnt overlock HBM, that reported hertz changes but its not doing anything
Meaning I can set it to 2000MHz and its still stable

So, I have been already reading this thread through few times to find out how yours turns out to be, but, I figured to just write it here


Actually, trying to deal with my Fury’s eccentric behavior ended up being more sane when I tried to stability test with Overwatch, then Witcher 3, and finally with Warframe, and that ended up being the way I’ll also do these memory sticks

Whats worst about eccentric behavior is that all that waiting and testing just ends up is crashing Witcher 3 around 1 hour mark, and then it turns out that all the past tests are also crashing with this game, and thats like couple weeks of tinkering time wasted to Valleying :man_shrugging:t2:

:roll_eyes:

Fuck that, 10s Warframe cooking and its real crash, tick down and its crashing to desktop after few hours or stable, and the next tick is going to be stable for months


#7

I certainly understand what you mean; stress-testing is a really pain in the ass, because you have to find something that is realistic, but also something that is repeatable. Finding a balance is difficult.

For the sake of this thread, I am sticking to 3dMark, if only for the consistency and repeatability. However, I have tinkered with Furmark, and I have found that I can get to 1100 “stable” with +0.006 V offset, and +20% power limit. Which one is more accurate? At this point in time, I am not sure, but I have both profiles saved, and once I find the extreme limits, I plan on testing out gaming.

That said, I am running stress-tests now, so I plan on posting some updated results.


#8

Alas, after much toil and trouble, it would seem that + 0.030 V for the voltage offset seems to be the upper limit that the VRM cooling solution can handle - and additional voltage will be needed to increase beyond 1100.

A clock rate of 1100 seems to be the upper limit that I can achieve with the present cooling solution.
(Reminder: present cooling solution.)

Similarly, it would seem that the only program I can use to modify the HBM clockrates is Sapphire’s Trixx software - which still carries the bug of locking the core clock rate whenever I press any button.

Presently, I have just been playing games at 1100, just to see if I can encounter any real-world instability. This project is all but done, but I shall spend part of tomorrow research the Trixx bug, as well as maybe cracking open my GPU and replacing the TIM for both the GPU/Memory and the VRM. I have aftermarket thermal compound, and thermal pads, for just the occasion.


#9

CAN THE FURY BE CONTAINED?

The answer is, both, yes and no.

… And I took pictures! Not professional, by any means. My workbench was my kitchen counter.


This first picture contained the tools of my trade: some Arctic Thermal Pads, and Prolimatech PK3. This is some of the better stuff you can get on the market.


In this second picture, you can see how the factory application is basically gooped on. I idly wonder if that much thermal compound might be applied so thickly, that it might be acting as an insulator? The VRM pads are spot-on, though.

In this last picture, you can see how hard the thermal pads of the VRM were. (This is not, by necessity, a bad thing.)

In the background, you can see the nineteen (19) count of screws, in three different sets, used to keep the PCB, backplate, and heatsinks melded together. The first two sets, four each, were the primary and second sets for the GPU to the heatsink. The remaining eleven (11) were to secure the backplate to PCB.


I am proud to announce that the changes in aftermarket cooling had… minimal effect for overclocking. A let down for all my toil and trouble, but great news for anyone who had not gone through all this trouble in the quest of overclocking. (In other words: Great job, Sapphire!)
Usually, I have no issues investing in aftermarket components. Manufacturers often try to save pennies by using things that can just “get by” and “be good enough” and it has happened to be often enough that swapping out thermal compounds has granted me an edge for overclocking. However, in this case, while I did see a slight improvement for temperatures, it had not practical effect.

So what did we learn?

Well, I guess, objectively, we learned that no matter what quality of components you use, you’re always at the mercy of the silicon lottery, and your ability to cool your components so they can continue functioning. Once your overclocking potential has been hit, the benefit that quality components grant you is a continued lifespan - not always something we think about, until something dies on us.

Personally, this test has reinforced my belief that it can we worth it to spend some extra money on a proven, robust product - especially if you plan on doing silly things like this, but, more importantly, for the piece of mind that you don’t have to do silly things like this.


#10

you should keep in mind
HBM Only Supports 500.00/545.45/600.00/666.66MHz And Rounds To The Nearest Step If You Set Another Memory Clock

you can have more luck pushing higher clocks with less voltage. If you manage to cool down your VRM well you may get better efficiency/stability overall.


#11

I think that was stated before but… kudos for the effort. I am about due to tear down my nitro furies and re-tim them as I have notice some creep. Thanks for the pics. My personal opinion is underclocking is where it is at with the Furys.


#12

Huh, I’m able to get 1100 MHz on the core clock without having to add any extra voltage. Maybe I’m lucky.


#13

I’m still waiting if HBM OC works or doesnt work, in other words, can you make HBM clocks crash


#14

Alas!

Sapphire Trixx is the only utility I can currently use to modify the memory clock rates, but whenever I hit “apply” the core clock gets reduced to, and stuck at, 300mhz. The modest time I have spent researching this tells me that it may be Wattman interfering with the program. So, I may try doing a clean re-install sometime, and seeing if I can forego Wattman, but another part of me is willing to let sleeping dogs lie.

Meanwhile, the heat and humidity here, where I live, is ridiculous, and I am waiting for it to cool down some before I try overclocking/undervolting more - if only for my own comfort. I’m not holding out for anything impressive, though, because I ran into some instability problems at -0.048 V offset at 1100mhz. We’ll see if controlling the heat has any effect, though.


#15

This is how the thread ends - not with a bang, but a whimper.

Tests with even -0.024V has issues when trying to run at 1100mhz. My chip just does not do well with the higher clock rates.

However, where trying to obtain raw power has failed, chasing efficiency has quite readily, and handedly, yielded fruit. I’ve been running stress tests, and it seems I can run the stock 1050mhz with an offset of -0.092V.

Madness. Complete madness.