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

[Bo_Fox]

I have shown that those numbers don't make sense. Nvidia cards still should be performing better then they do. With the numbers from a GTX285, they are not using the MUL to it's full potential, one of the things that was supposedly "fixed" from G80. It is not performing at the 1.06TFlop(MADD+MUL) level nor at the 708GFlop(MADD) level but an inbetween. Basically the MUL is only being exploited ~45% of the time.

Sure, it is loading the cards but it is not exploiting all the shaders that the G200 has to offer.

People are complaining that they cannot fully utilize their RV770 cards, since some have to be downclocked to be stable in OCCT.
The flipside is that you are also not getting full advantage of all that G200 has to offer with the results of OCCT.
When using a G200 you are only receiving ~82% of the performance.

1.06TFlop for MADD+MUL still isnt as good as 1.2TFLOP for 4870 cards. Please elaborate.

[LordEC911]

1.06TFlop for MADD+MUL still isnt as good as 1.2TFLOP for 4870 cards. Please elaborate.

Here...

I wasn't trying to prove a point, I was simply answering a question.
I can see where he might have been headed by asking that question though.

4890 is what ~10-15% behind a GTX285 with both at stock on average in "normal" games and apps?
Yet with OCCT, the 4890 is ~56% faster, using the 83FPS vs 53FPS.
However this app is programmed it stresses every part of the chip to the max, or at least quite a bit more than other "normal" apps/games.

Also none of the numbers, i.e. FPP, seem to add up.
4890@850mhz= 1.36Tflops
GTX285@1476mhz= 1.06Tflops(MADD+MUL), .708Tflops(MADD)

1.36/1.06= 1.28x greater (1/2 the FPS difference)
1.36/.708= 1.92x greater (amusing since it doesn't mean anything but = largons power draw increase)

Simply using max theorectical FPP is not an accurate way to estimate performance but in this case it seems to be related. Since this app has been said to use simple shaders to completely load the ALUs, you could come to the conclusion that the MUL is only being used ~45% of the time.

Basically, the way this app is programmed it is able to use the 4890's architecture to the max and seems to not fully load Nvidia cards, persay.

Edit- Anyone know what the stock volts for a GTX280 is under load? 1.3-1.4v?

[Bo_Fox]

Here...

Perhaps more of Nvidia's processing power are "reserved" for PhysX calculations. I wouldnt be surprised if many of those transistors are wasted anyways on 32 ROPs, 512-bit memory interface, and espeically PhysX.

After all, nothing ever pushes closer than 99% of the theoretical maximum. Some of the older cards were so inefficient, like G92 cards, in only getting within 75-85% of the theoretical maximum for fillrate tests.

[W1zzard]

Perhaps more of Nvidia's processing power are "reserved" for PhysX calculations.

according to nvidia nothing is reserved. when an app requests cuda, the driver will set some shaders aside for that task for the lifetime of that cuda application.

[cx-ray]

I wasn't trying to prove a point, I was simply answering a question.
I can see where he might have been headed by asking that question though.

4890 is what ~10-15% behind a GTX285 with both at stock on average in "normal" games and apps?
Yet with OCCT, the 4890 is ~56% faster, using the 83FPS vs 53FPS.
However this app is programmed it stresses every part of the chip to the max, or at least quite a bit more than other "normal" apps/games.

Also none of the numbers, i.e. FPP, seem to add up.
4890@850mhz= 1.36Tflops
GTX285@1476mhz= 1.06Tflops(MADD+MUL), .708Tflops(MADD)

1.36/1.06= 1.28x greater (1/2 the FPS difference)
1.36/.708= 1.92x greater (amusing since it doesn't mean anything but = largons power draw increase)

Simply using max theorectical FPP is not an accurate way to estimate performance but in this case it seems to be related. Since this app has been said to use simple shaders to completely load the ALUs, you could come to the conclusion that the MUL is only being used ~45% of the time.

Basically, the way this app is programmed it is able to use the 4890's architecture to the max and seems to not fully load Nvidia cards, persay.

Edit- Anyone know what the stock volts for a GTX280 is under load? 1.3-1.4v?

I just ran the GPU test on a pair of GTX 280s SLI for about 10 mins (clocks 712/1512/1242 - voltage 1.185). Full screen 1920x1200 settings. Max current draw was just below 30A on both cards. Max frames per second 144. Both cards are water cooled. Temps of GPU0 max 56C and GPU1 53C. Ambient temp was ~23.8C. During gaming temps usually top out at 15C above ambient.

[zalbard]

I just ran the GPU test on a pair of GTX 280s SLI for about 10 mins (clocks 712/1512/1242 - voltage 1.185). Full screen 1920x1200 settings. Max current draw was just below 30A on both cards. Max frames per second 144. Both cards are water cooled. Temps of GPU0 max 56C and GPU1 53C. Ambient temp was ~23.8C. During gaming temps usually top out at 15C above ambient.

Keep in mind different settings should be used in order to fully load Nvidia chips with this test. They were posted earlier in this thread. Also was VSync off?

[cx-ray]

Keep in mind different settings should be used in order to fully load Nvidia chips with this test. They were posted earlier in this thread. Also was VSync off?

Not sure where to find the VSync off or even if it matters as the fps is already higher than the standard 60hz of an LCD. So, it doesn't look like it's sync locked.

New settings I just tried VSync forced of in the Global Settings tab of my driver (didn't see a difference).

13:27 min. with Shader Complexity settings 3, FPS ~83, GPU temps 56C and 53C, current draw still about 30A. Ambient temp ~23.4C.

I tried to search for other settings in this thread. The above was the closest I could find for Nvidia cards. If you know something better let me know and I'll test again.