OCCT 3.1.0 shows HD4870/4890 design flaw - they can't handle the new GPU test !

[largon]

SparkyJJO,
Maybe your reference Visiontek has a revised bios with a higher OCP trigger?

[Sparky]

SparkyJJO,
Maybe your reference Visiontek has a revised bios with a higher OCP trigger?

Doubt it. I got mine only a couple weeks after release
(For $185 )

[lyserg]

i Dont know if it helps u but here :

This was done with a newer Version of Powercolor HD 4870 1GB PCS
800/925 (stock on new revision is 780/925)
BIOS 02/24/09 21:07
ATOMBIOSBK-ATI VER011.013.000.002.000000

Like u said before ,no Blackscreen

But i have the same Powercolor with old Revision,i´ll test it in the next half an hour.

Now this was the older Revision
800/925 (stock)
BIOS 09/04/08 04:35
ATOMBIOSBK-ATI VER011.010.000.001.029909
Wekiva RV770 B50702 Board

Instant Blackscreen after the Donut comes out

[Tetedeiench]

Same like some others I believe the heading is little sensational and even misleading, since OCCT is sure not the only GPU benchmark whether new or not.

Now we have basically two arguments here going on and on.

First even if this can't be duplicated in ANY game, some people including Tetedejench argue MAYBE one day this problem will show in future games. We all know sooner or later all Graphic cards are going to be obsolite. Since the game developers need to sell many games it usually doesn't happen too fast.

The other argument is, nobody plays OCCT so who cares and I kind off agree with that since I don't have any problem with the card at all. Looks like mainly Nvidia guys have problem with the card even if they don't have any.

Now I did run the OCCT shader 3 bench with NO error check and it crashed immediately so it could not have anything to do with overheating.
It crashed so badly that i could not get the Vista 64 Appearance Option till second reboot, so I was telling myself this OCCT really sucks.

Anyway I tried too more settings.

First with Shader 1, NO error check and it run OK. Here is the graph, 1 minute idle, 1 minute test and 4 minute idle.


Than I tried to run the Shader 3 with the error check ON, same times as above and it did not crush at all. The graph looks the same as the above one except the Error Check. The test went in full screen but the "furry wheel" was not moving yet the test looks the same including FPS reported. This Graph reports "Test Completed without errors" which is not right.
So the simple application also has errors. Here is the graph.


The test was run with 4870 X2 but I also tried without the crossfire with same results.

PS English is not my first language so don't get too critical about that.

Simulated errors on the 3d app by scrambling the reference image in the 3d test. Here is the result :


See the error count ?

Everything is fine there. The error count is not there to count the number of crash of your card, but the number of computing errors your card produced. Just as any other stability checker.

[Tetedeiench]

For the one who requested it :

GTX285, stock frequencies, Shader Complexity 3, 1680x1050 FullScreen Mode :


Sorry for the french header, forgot to remove it.

had about 53FPS average.

[villa1n]

I didn't ignore your post. It's hard to answer everybody

I actually asked for more information :

• Did you use Shader complexity 3 ?
• Can you take a picture of the VRM of your cards ? here is an example : http://mangafranceworld.free.fr/Divers/DSC00550.JPG
• Can you redo the same test and monitor the VRMA just as i did using RivaTuner, and post the screenshot ? The idea is to see if your card stays below, or higher than 82A on the VrmA. Another user did it on Page 10

If your card withstand the test, either the VRM stage is not the reference one (i.e. 4-stage, and not 3-stage), or you forgot to check the fullscreen checkbox Or it's something new !

I m pretty sure mine are both non reference.

My powercolor/gecube 4870 looks like this
gecube
and this is the sapphire one i have
sapphire

I used complexity 3 on it, fullscreen, but i heard my fans kick up to 100% instantly. Not sure if i want to run it again, as I dont want to hear a popping sound in one of my cards, cant afford to replace them at the moment ^^. I think both these cards are 4 phase though.

Just tried to get the volt readings via riva. Might be because i m on windows 7.. but i dont get the option for any Voltage readings.. just fanspeeds and temps. Riva says the Volterra regulators are only on the reference design, so mine must both not be reference.

[Tetedeiench]

I m pretty sure mine are both non reference.

My powercolor/gecube 4870 looks like this
gecube
and this is the sapphire one i have
sapphire

I used complexity 3 on it, fullscreen, but i heard my fans kick up to 100% instantly. Not sure if i want to run it again, as I dont want to hear a popping sound in one of my cards, cant afford to replace them at the moment ^^. I think both these cards are 4 phase though.

Just tried to get the volt readings via riva. Might be because i m on windows 7.. but i dont get the option for any Voltage readings.. just fanspeeds and temps. Riva says the Volterra regulators are only on the reference design, so mine must both not be reference.

I wonder if we should start talking about 3-phase and 4-phase VRMs at that point

We had a sapphire that could withstand the test, and it is indeed equipped with a 4-phase VRM. I even have a video of that :
http://www.snapdrive.net/files/235474/P1030851.avi
52°C to 100°C in 13 seconds

Might be the very same card as yours. 4 phase VRM indeed. Here is a pic of the card VRM, pried apart : 4 phase.

[initialised]

I beg to differ, I blame the test rig.

A while back I found that ASUS 4850s simply would not work on P43 (P5QL-Pro & P43 Neo) motherboards. Other brands were fine in the same systems. The ASUS cards worked fine in other systems.

Yesterday I found the same with XFX 4850 1GB on 790FX DDR2 boards (K9A2 Platinum, DFI 790FX M2RSH, M3A78T. Likewise, swapping them out for Powercolor cards did the trick and the XFX cards tested fine on a different platform.

The failure mode is the same as you described.

In other systems with the same failure mode increasing the PCIe bus voltage to 1.6V can help.

So I'm thinking that it is the PCI-E slot that cannot deliver the power asked of it or detects over-current condition and shuts down the graphics card on certain platforms.

Also cards from certain vendors don't work properly with certain brands of memory, I seem to remember OC'd 8800GTS512 cards failing repeatedly with certain brands of cheap DDR2, change the brand of RAM or GPU and it would work fine.

How many cards are affected ?
Right now, we've successfully crashed about 10 different cards using this test, using alot of different power supplies (ranging from 550W Antec to 1500W ToughPower (!!!). We had Seasonic, Corsairs... etc).

Since you don't mention changes to CPU, RAM or motherboard or list the test platform(s) you used it is possible that part of the test platform is the problem. If it repeats with all reference 48x0 cards at stock clocks on every current platform then you might be on to something, but then you need to vary driver versions operating systems and your own OCCT build. Basically you need to expand your test matrix before you can jump to your conclusions.

BTW it's my job to stress test systems based on pretty much all the current platforms with most available graphics cards, does that qualify me as a hardware guru?

[Solus Corvus]

What mod is needed to raise the OCP limit on a 4890? Is it the same as for the 4870? If nobody else wants to test that I will when I get my new card.

[Tetedeiench]

I beg to differ, I blame the test rig.

A while back I found that ASUS 4850s simply would not work on P43 (P5QL-Pro & P43 Neo) motherboards. Other brands were fine in the same systems. The ASUS cards worked fine in other systems.

Yesterday I found the same with XFX 4850 1GB on 790FX DDR2 boards (K9A2 Platinum, DFI 790FX M2RSH, M3A78T. Likewise, swapping them out for Powercolor cards did the trick and the XFX cards tested fine on a different platform.

The failure mode is the same as you described.

In other systems with the same failure mode increasing the PCIe bus voltage to 1.6V can help.

So I'm thinking that it is the PCI-E slot that cannot deliver the power asked of it or detects over-current condition and shuts down the graphics card on certain platforms.

Also cards from certain vendors don't work properly with certain brands of memory, I seem to remember OC'd 8800GTS512 cards failing repeatedly with certain brands of cheap DDR2, change the brand of RAM or GPU and it would work fine.

If it repeats with all reference 48x0 cards at stock clocks on every current platform then you might be on to something. You need to expand your test matrix before you can jump to your conclusions.

BTW it's my job to stress test systems based on pretty much all the current platforms with most available graphics cards, does that qualify me as a hardware guru?

we had alot of different configurations, from i7 configs to AMD configs showing the problem. Brands were Asus, Gigabyte, MSI (i'm taking this list from memory). So i'd say that unless they all messed up implementing the PCI-E specs, and we are onto something else.

That's still something possible though. I'd think it unlikely, but possible.

But if your hypothesis is true, how come the VRM-overpowered cards (4-phase) are running flawlessly ? That doesn't make any sense if the problem comes from the PCI-E port. No ? Those cards could go up to 108A... how come others are limited to 82A ?

[Nightstar]

Bravo Tetedeiench I've grown quite fond of your OCCT. Don't let the fanboys and shills shout you down.

The Asus non-reference 4870 1GB has no issues.
My Visiontek reference design v1 (there's a later reference design as well IIRC) 4870 512MB has no issues. Which strikes me as odd that any others based off the reference design would, what makes mine different?

I've been told that GDDR5 stuff is power hungry, could be your 512MB card draws less than a 1024MB reference design card and doesn't trigger OCP.

If the card can be made to draw current in excess of either its supply capability or its current protection threshold using nothing other than normal API commands, there's a design problem.

Or a manufacturing problem. Nonetheless /agree wholeheartedly

It would be even funnier and sad if this will result in deterring people from buying Radeons.

Nah, here is the sort of thing that would deter me from buying Radeons:

"Your limited testing is flawed. If the VDDC's overheat they will shut down the card to protect it. If you dont want to have a crash when testing artificial means such as this stability benchmark then increase the fan speed to cool the VDDC's. Obviously underclocking acomplishes a similar goal as they dont get as hot. No games are affected just artificial stability tests."

I was considering a 4890 however I won't abide this kind of BS and censorship. I certainly won't encourage such with my hard earned $.

This has been a very revealing thread. I've learned a lot about individuals that compose the XS community

[LordEC911]

For the one who requested it :

GTX285, stock frequencies, Shader Complexity 3, 1680x1050 FullScreen Mode

had about 53FPS average.

Sooo... the Nvidia cards perform worse than the AMD/ATi cards?
Doesn't the AMD/ATi card also outperform the Nvidia cards in Furmark?

So how is this not a power virus?
You create an app that creates an unrealistic load on the card, causes the card to draw an unrealistic amount of power and then blame the card?

[Sparky]

Does anyone know how many phases the visiontek 512MB 4870 has? I really don't feel like pulling the sinks tubes and sinks off my 4870 to find out.

[xlazyx]

Ati realy messed up here.

[G0ldBr1ck]

Lets see how long this one lasts...heh

http://forums.amd.com/game/messagevi...&enterthread=y

EDIT: Gone all ready, I was polite as one can get and they still deleted my question. WTF

EDIT 2: Oh, they banned me to. what BS

[largon]

Does anyone know how many phases the visiontek 512MB 4870 has? (...)

You can tell by look at the backside of the card. If the marked caps are missing it's 3 phase, if they are there it's 4 phase.

Sooo... the Nvidia cards perform worse than the AMD/ATi cards?

Yep.
A downclocked HD4890 (850/850) keeps >80FPS constant:

(OCCT doesn't support temps monitoring on HD4800s so the second line is 5V)

[ryan92084]

ran for 5 minutes on an original reference 4870 1gb (visiontek or HiS i forgot) with the latest win7 drivers. I stopped after 5 minutes because it didn't look like it was going to implode but i didn't like not being able to see the vddc temps while fullscreened.

card is watercooled (not a FC block) while the mem and VDDC are passively cooled with mildly modified iandh sinks. i cool both my cpu and gpu on the same loop with a single 240 radiator.


room = 25C
card clocks= stock

gpu idle = 40C
VDDC idle = 55, 57, 56C

gpu max =48C
VDDC max =136, 141, 137C

all temps were pretty normalized at this point. i've never had my vddc temps go that high before (furmark only put them up to 109 or so after 10 minutes and while gaming they never get anywhere even close to that) but this is an exceptionally warm day.

i don't really see how this is new or special. its the same issue furmark causes just more so.

[initialised]

we had alot of different configurations, from i7 configs to AMD configs showing the problem. Brands were Asus, Gigabyte, MSI (i'm taking this list from memory). So i'd say that unless they all messed up implementing the PCI-E specs, and we are onto something else.

That's still something possible though. I'd think it unlikely, but possible.

But if your hypothesis is true, how come the VRM-overpowered cards (4-phase) are running flawlessly ? That doesn't make any sense if the problem comes from the PCI-E port. No ? Those cards could go up to 108A... how come others are limited to 82A ?

Maybe when the VRM's hit the amperage limit the card tries to increase power draw from the PCIe resulting in an over-current condition and shuts down the port, that still puts the alleged fault with the card not the board. Overvolting the PCIe bus might get around this on some systems.

Try using the three data points makes a line in two dimensional space approach to formulate a test:

3 3-phase cards from 3 vendors
3 4-phase cards from 3 vendors
3 different X58 boards
3 different Dragon boards
3 different P45 DDR3 boards
3 different sets of DDR3 RAM
3 different OSes Win-7, Vista, XP
3 different DX9 builds

Test each card on each platform with each set of RAM with a fresh install of each OS and each DX9 build. If the 3-Phase cards all fail and all the 4-Phase cards all pass in all configurations. Then you can be fairly sure that the 3-Phase cards are incompatible with your test.

Then the question becomes 'is it your test software, DirectX or the cards that is at fault?'

If you rewrite the test from scratch in OpenGL or DX10(.1) will the fault still occur?

Does data from GPU-z corroborate what you have found with RivaTuner?

Does data from multi-meters corroborate with software monitoring apps?

I'm not saying that your findings are wrong, I doubt that they are but convincing ATi and their board partners that 3-Phase power is not enough for will probably take the level of scientific rigour I've proposed above.

Even then is it relevant if there are no commercial (let alone AAA) games that are capable of stressing the hardware to this level while retaining playable framerates?

[LiquidReactor]

Fff.....I ran my HD 4890 oc'ed to 1.033 vGpu 1.39V in OCCT without any problems. You guys probably are undervolting the pcie slots or using crappy psus.

[[XC] Lead Head]

The way I see is like buying a brand new 600 HP Sports car, only to find out find out when you plant the pedal to the floor, the fuel pump can't pump enough gas to the engine, and then the engine dies/loses power. Realistically you will NEVER need all 600 of that HP in normal driving, so would you then claim that nothing is wrong with your car's design because you will never really need to go wide open? That the engine died just because you pushed the gas pedal to far down then it ever normally would, even though the company that built the car lets the pedal go that far down in the first place? I don't think anyone would say there was not a problem with their car.

This is the way I am seeing it.

Yes the test is entirely un realistic, but the fact that it can still "over load" part of the card is not good. AMD and Intel release specs for their CPUs that the voltage regulation of a motherboard has to have the ability to meet a certain current value for their most power hungry CPU that could potentially use said socket. Apparently ATI does not.

[nr4]

This IS a serious and very likely (though certainly unacceptable design flaw). Look at any inductor spec sheet and you will see a few numbers that describe the way the inductor behaves. As load is applied, DC bias rises, and inductance falls the inductor reaches a current threshold which causes the temperature to rise by about 40C (Heating current or Iheat) which usually coincides with or is not far from the inductor's Irated. In the area of the Inductance V DC Bias (current) curve within specifications, inductance remains relatively the same. When it reaches Iheat Inductance begins to fall rapidly until current hits the third important number, saturation current or Isat, at which point inductance drops %20, i forget whether it's an asymptotic behaviour, higher PCB and ambient temperatures will cause this to occur at lower current. Therefore, there is a threshold current where the buck converter will no longer be able to meet power demand, i mean that's rather intuitive.

I have to say that Slovnaft gave a very good answer on this matter. He explained very well how the inductor properties are changing with current and temperature.
This is one of the most plauzible explanation of the phenomenon.
An inductor of 3 phase vrm is faced to much more heat and power in time than 4 phase vrm.

Also i have to congratulate Tetedeiench for his work in general, not particulary for this discovery.

But in the end i would say that this problem should concern the videocard makers that in the future should give more importance to the vrm and stop building cheap ones.

HD4870&HD4890 are very good video cards and wont present problems in any game or app, years from now.

[LiquidReactor]

Would be nice if op could provide more evidence

He so far tested 2(?) HD 4870 cards and one of them was defective (?)

All this without any serious pics or any screenshots of the tests etc...besides the rivatuner shot of HD 48xx series.
Heck, op probably has an overvolted hd 4850 that he is tearing to bits with this new test of his.

Multiple HD 4870 owners in this thread tried his test and their cards passed.

Has op even posted proof that he actually owns a HD 4870?

[Slovnaft]

guys. it's inductor saturation, cmon. rv770 and 790 are pretty hot chips, volterra pwm is an incredibly hot solution to begin with, and heat reduces Isat threshold. I mean, it fits all the symptoms.

@Tetedeiench, can you try monitoring the surface temp. and inductance of the inductors as this 'phenomenon' occurs?

[Praxis1452]

Sooo... the Nvidia cards perform worse than the AMD/ATi cards?
Doesn't the AMD/ATi card also outperform the Nvidia cards in Furmark?

So how is this not a power virus?
You create an app that creates an unrealistic load on the card, causes the card to draw an unrealistic amount of power and then blame the card?

Get Out!

http://en.wikipedia.org/wiki/Power_virus "A power virus is a malicious computer program that executes specific machine code in order to reach the maximum CPU power dissipation (thermal energy output for the central processing unit). Computer cooling apparatus are designed to dissipate power up to the Thermal Design Power, rather than maximum power, and a power virus could cause the system to overheat; if it does not have logic to stop the processor, this may cause permanent damage."

OCCT is not malicious code.
The wiki entry even makes a distinction between a power virus and a stability program in that a stability program is directly under user control. And OCCT is.
It has not been proven that this causes the GPU to overheat, rather what people suspect is that OCP is being reached, or degradation of an inductor has occurred shutting down the graphics card. Thereby heat is not the main issue.
^^ this is the hypothesis under question though more broadly it is why this has ocurred.

You have not contributed a single piece of data, nor have you augmented the hypothesis and analyzed the results. Your only conclusion is that this is caused by temperature(because you cited power virus), which is wrong.

You're useless, and blatantly lying so you're absolutely worthless. Destroying you is fun.

[Praxis1452]

You guys are getting trolled hard by op.

He so far tested 2(?) HD 4870 cards and one of them was defective (?)

All this without any serious pics or any screenshots of the tests etc...besides the rivatuner shot of HD 48xx series.
Heck, op probably has an overvolted hd 4850 that he is tearing to bits with this new test of his.

Multiple HD 4870 owners in this thread tried his test and their cards passed.

Has op even posted proof that he actually owns a HD 4870?

Multiple owners also failed. lol nice way of ignoring the data.