Jump to content
HWBOT Community Forums

Intresting find about Fury X performance with software voltage control


buildzoid

Recommended Posts

So I kept getting negative scaling with my Fury Xs in benchmarks and so I decided to do some formal investigating.

 

This is what I got:

attachment.php?attachmentid=3910&stc=1&d=1455582043

 

The numbers are from Unigine and as you can clearly see raising core voltage in Trixx is causing performance to fall of a cliff. Power limit doesn't have any impact on this. I tried to fix this by using the 350W Sapphire Fury Tri-X OC BIOS. That didn't do anything so I modded the BIOS for 768W. That still didn't help.

 

The performance hit from voltage is so bad that 1175mhz core clock at +100mv gives the same FPS as 1125mhz core clock.

 

I still need to check if this happens with hard mods for voltage. If it does I'll probably try a hard power mod too.

Link to comment
Share on other sites

Interesting find! It may be worth your time to plot temperatures as well in case that's what is causing the throttling.

 

Core never went over 46C during testing. I think when I started testing I was at 43C. Mind you after each set of results I went and tested that stock clocks still gave 274FPS. The difference in performance is instantaneous. As soon as GPU-z picked up on the raised Vcore the FPS was down.

 

I tested with Trixx intially but Afterburner also does the same.

 

 

If I can fix this I see an easy 19-20K graphics score in 3Dmark Firestrike. Because stock clocks does 17.8K and the cards seem capable of core clocks between 1200 and 1250mz with enough voltage.

Edited by buildzoid
Link to comment
Share on other sites

Does % performance loss match (well.... mirror) % voltage increase?

 

If so, I call shenanigans.

 

 

(really nice find, though! .....and thankyou for posting it up!)

 

Looks exponential to me: gtnoGKO.png

 

The only thing that goes up exponentially with voltage is power draw which is why I think it's a power management problem.

Link to comment
Share on other sites

The one thing missing from your test data is the result of increased MHz and increased voltage. THAT will get the power consumption up :D

 

Using 1.175v // 274fps as a baseline, a 17% increase in voltage takes ~~6.2% off the score. The other data points show less.

 

My theory of shenanigans is, so far, inaccurate :D

 

 

 

Power consumption shouldn't increase exponentially with voltage.

Link to comment
Share on other sites

The one thing missing from your test data is the result of increased MHz and increased voltage. THAT will get the power consumption up :D

 

Using 1.175v // 274fps as a baseline, a 17% increase in voltage takes ~~6.2% off the score. The other data points show less.

 

My theory of shenanigans is, so far, inaccurate :D

 

 

 

Power consumption shouldn't increase exponentially with voltage.

 

P = I * V

 

I = V/R

 

P = V * V / R

 

That's the theoretical proof for power draw being exponential with voltage. I also have a ton of real world test data supporting that theory.

 

I'll do more testing but so far I think Power Tune is completely broken and it's doing some kind of micro throttling.

 

Also I think I've managed to trip OPP at 35A on the PCI-e 8 pin. I'm gonna hack Trixx for even more Vcore and see if I can get it to happen by just raising core voltage really high. Just to make sure that crash was indeed OPP and not an unstable core clock. (I'm so gonna end up breaking my Fury X)

Link to comment
Share on other sites

You can prove that it's not exponential as well. It's the first thing that you typed!

 

All of us have a ton of real world data on how it's not exponential, thanks to OC

 

If I increase vCore by 60% without changing MHz... what % does the power consumption increase by?

 

If I increase vCore by 60%, without changing MHz... AND hold the temperature steady...what % does the power consumption increase by?

 

:D

We can argue this either way.

 

 

EDIT...

 

I guess our typical voltage testing range is too small to show the exponent. Take a small enough interval and everything is a straight line......

 

But P=IV is a linear.

Edited by K404
Link to comment
Share on other sites

You can prove that it's not exponential as well. It's the first thing that you typed!

 

All of us have a ton of real world data on how it's not exponential, thanks to OC

 

If I increase vCore by 60% without changing MHz... what % does the power consumption increase by?

 

If I increase vCore by 60%, without changing MHz... AND hold the temperature steady...what % does the power consumption increase by?

 

:D

We can argue this either way.

 

The math for a single transistor is simple atleast, power consumption is linear as a function of frequency and squared for voltage if you dont take other effects into calculation.

 

A GPU sure is more advanced then a single transistor so it wont be a simple square value but its no suprise that the effect of vCore wont be linear atleast

Link to comment
Share on other sites

MOAR DATA:

 

O9PVEDy.png

 

I dropped core clock so I could test at negative voltages. The power draw number is current on the 8 pin PCI-e connectors.

So behaviour wise this looks pretty much identical to the results at 1050 core clock. The power draw readings also don't reveal any useful trends other than the fact that power draw scaling is mostly linear. Now this goes against all the data I have from measuring power draw on CPUs. However the drop in FPS means that the GPU is doing less work so I'm still going to stick with my theory of Power Tune messing with something which results in the linear power draw scaling and drop in FPS.

 

I do have an idea for how to prove that power draw is exponential with voltage. I'll just run the card at 500mhz core clock and test it with core voltages ranging from .95V to 1.175V since there power tune doesn't seem to be intervening.

 

Also I tested if 35A is the OPP. It isn't I ran +252mv with 1050 core and the card ran fine for several seconds which if OPP was at 35A would have been a trip.

Link to comment
Share on other sites

Is core clock = turbo clock? If yes maybe we need disable turbo clock. Maybe turbo cause this.

 

AMD GPUs don't work at all like Nvidia. Nvidia cards automatically clock up but AMD cards are supposed to run at what ever clock you set them at unless they are overheating or hitting the power limit.

Link to comment
Share on other sites

Ok so I did another round of testing. I didn't do a write up since it was a really basic check.

 

I underclocked the GPU down to 500mhz. I set core voltage to about 1.4V surprise surprise no FPS drop. With these settings the card was pulling only 222W. Stock power draw for comparison is 312W. Mind you I chose a specific view in Heaven which causes AFAIK the highest possible power draw.

 

So the conclusion is that the power management is intervening based on power/current limit because when running 1.4V with a 222W power draw everything was fine but running 1.4V with 420W power draw causes the massive FPS drop.

 

From my testing with modded BIOSs ranging from a 4096 core 350W Fury Tri-X BIOS from Sapphire to my own 65000W BIOS I've also come to the conclusion that the BIOS doesn't have any impact on the amount of throttling occurring. So a physical power mod should work however physical power mods on the IR3567B are risky because the current sensing is per VRM phase and is used to load balance the phases. There is a chance a Physical Vcore mod might also work but I can't check because ATM I don't have what I need to do a hard mod. I've tried BIOS modding however I probably have no idea how Fury X BIOS works.

 

If some one could test with their own Fury X for the same effect that would be great.

Edited by buildzoid
Link to comment
Share on other sites

  • 3 weeks later...

Viewing your data and what I've picked up about "PowerTune" I would concur you're seeing FPS drop due to that.

 

ASIC Profiling is going on in Fiji like Hawaii, ASIC profiling is used so a ROM does not have to be tailored to set VID per GPU. Besides other GPU properties it is based also on LeakageID.

 

Now why LeakageID matters in the context of this thread IMO is:-

 

Also it must be noted that GPUs with low and high leakage properties have different voltage under load eventhou the voltage would be set to the same level. A low leakage GPU might draw 150A of current in load while a high leakage part can draw 170A or even higher. Since the load-line resistance (RLL) is fixed instead of being dynamic based on the leakage current, the high leakage GPU will have a greater load-line effect (voltage drop).

 

E.G.

 

Low Leakage GPU: VDDC = 1.20000V, IDDMax = 150.0A, RLL 0.64mOhm

 

Idle (~ 5A IDD) = 1.20000 - (((1/10000)*6.4) * 5) == 1.1968V

Load (~ 150A IDD) = 1.20000 - (((1/10000)*6.4) * 150) == 1.104V

 

High Leakage GPU: VDDC = 1.20000V, IDDMax = 170.0A, RLL 0.64mOhm

 

Idle (~ 5A IDD) = 1.20000 - (((1/10000)*6.4) * 5) == 1.1968V

Load (~ 170A IDD) = 1.20000 - (((1/10000)*6.4) * 170) == 1.0912V

 

Quote from link

 

I believe I currently have a low leakage Fury Tri-X card (Fiji Pro), it has a VID of 1.243V for DPM 7. Viewing GoldenDB data in PowerPlay 1.250V is MAX VID a low leakage ASIC would have under EVV for stock setup.

 

Only problem using VID to determine leakage is even 2 GPUs with identical LeakageID can have differing VID due to the other properties deeming it say good or bad ASIC within LeakageID bands. I can not elaborate further as it has taken me some time to wrap my mind around this stuff from when doing Hawaii bios mod.

 

I will run similar test as you but don't have access to or know how to use electrical test equipment. Therefore my results would not have that data, could get some data via HWiNFO.

 

I also have a Fury X (Fiji XT), not installed it yet, I have another being delivered tomorrow, I'm hoping both have differing VID for DPM 7 = differing leakage to the Fury Tri-X. So if they draw differing amps / watts under load due to leakage, using same clocks it may show as a FPS variance.

 

Did you use FRAPS to get FPS data for Ungine Heaven or the in app results?

 

Even though we can edit the 3 PowerLimit values of TDC (A) / TDP (W) / MPDL (W) in PowerPlay of ROM, I reckon the driver or something else is limiting card. Just like you and I have seen we can not set higher than 1.3V VID for a DPM in PowerPlay via ROM. Only solution which is working is to add a VDDC only offset to VRM controller in ROM in conjunction with VID mod in PowerPlay.

 

I don't believe currently the "PowerTune" limit being encountered is due to also the VRM controller chip. For example values which I deem involved with OCP, etc in Fiji i2cdumps / VoltageObjectInfo of ROM are only slightly lower than Hawaii.

 

For anyone interested or wishing to get involved in Fiji Bios mod use this link.

Link to comment
Share on other sites

Firstly my apologies if double posting is not ok on forum :blush: .

 

Fitted the Fury X last night, modded the cooling profile so AIO maintains =<55C, also upped PL to 297A / 330W / 330W (stock 270/300/300).

 

qzwI7X4.jpg

 

So pretty much at stock clocks I'm near PL limits, currently as Martin Malik stated HWiNFO uses GPU's own telemetry info for values shown it's a clear indication of what "PowerTune" would use to assess when to apply limit,etc.

 

Using VID as rough estimate of leakage I have DPM7 VID@1.250V (low leakage ASIC). Through some registers / i2cdumps gained on OCN I know of a member with Fury X has DPM7 VID@1.231V and another has as low as DPM7 VID@1.225V.

 

So taking The Stilts info regarding Hawaii and GoldenDB VIDs in Fiji ROM I'd say the members with 1.231V & 1.225V will draw more A and hit "PowerTune" limit sooner / see effect more aggressively.

 

What is VID for DPM 7 of your Fiji on stock ROM / clocks?

Edited by gupsterg
Link to comment
Share on other sites

  • 4 months later...

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

×
×
  • Create New...