have you tested your overcloked 4870 with the fur rendering of GPU Caps Viewer? I'd like to know if FurMark and GPU Caps Viewer have the same behaviour.
FurMark is just an OpenGL app and has no low level code (I mean no assembly or direct access to the hardware). So this is certainly a radeon hardware issue (maybe the bios, I don't know...).
As per your request I have done some testing and here are the results.Things to know about the 4870
It is designed to Idle at 80C. It actually tries to stay there. You can run it in a 21C room with basic store bought floor fans blowing on it through the side of your open case and it will almost stop it's own fan and unlikely to drop below 77C. ATI engineers actually intended on not letting the card drop far below 80C. It's by design. Under gaming load it is cooler running than most factory overclocked 8800GT's that, as we all know, have a HUGE install base. In regular gaming sessions, my 2 8800GT's tested at different times hit 90C in this same case I use for the 4870. This was tested in the middle of winter so the room was 20C. The fan on the cooler used to spin something awefull in some games. The 4870 also incorperates a VPU recovery algorythm programmed into the card. If the card detects a potentially unsafe temperature in one of it's components, it shuts down the card. From what I understand from talking to ATI Beta testers, it's just over 100C somewhere. Sounds reasonable and a good idea to boot. Don't want to cook a card right?
A "stress test" should infact demostrate the card's design in these respects. On with the testing...Test scenario and background
This is a stock system. Intel board and Intel C2D 8400. All stock. As far as testing a 4870 goes, this is about the worst case scenario I can imagine. Firstly I run the Atnec Sonata III. All reviews of this case end with the reviewer saying "NOT recommended for overclocking as it does not move alot of air (hence the silence) and as a gaming rig, we strongly recommend you add a front 120MM fan to help cool your cards
". Needless to say this is a very hot case. Ironically this presents a good testbed for card stability.
Room ambient temps on the floor at the front of the case = 27C. (~ 82F)
Using current version of GPU Caps Viewer. 600x600 windowed res (seems to get the hottest)
No other 3rd part apps (ATT, Rivatuner, GPUz ect...) running or installed.
Cat 8.7 + Vista 32
4870 single card
Temps and Framerates read from top titlebar of GPU Cap Viewer app.Results with stock clocks - 750/900
Average FPS = 229
60 seconds - 96C
120 seconds - 97C with momentary touches of 98C
180 seconds - 99C
User ended test.Results with basic overclock - 780/1000
Average FPS = 240
60 seconds = 98C
120 seconds = 100C
180 seconds = 102C
Card failed and system blackscreened.
Temp at time of failure = 102CRE-TEST
It is 23C in the room. 4C cooler than the above testing. This is about 73F. Still warmer than room temperature. But a good time to test again.
300 seconds - 96C
User ended test.Conclusion
GPU Caps Viewer test was repeatable and predictable and the card behaved as per it's design and the environment it was tested in. 4870 needs airflow if you are going to overclock. 30C rooms are not good environements for this card if oyu overclock. At room temperature and proper airflow however, overclocking these is fine with stock fan speeds.
The overclock (given the nature of my crappy airflow and warm room temp) is generally solid. I can loop Devil May Cry 4 for hours (about the hottest test I can find) and the temps max out at 94C when the room was 30C to begin with. Most of the time the temp of the GPU would hover ~90C. In most other games it's in the mid to high 80's with this overclock. In the right case, or if I added a fan in the front of the case like I am supposed to, these temps would all drop ~5C. Especially at lower room temps.
My problem with Furmark in it's pure form (the furmark app and not running through GPU caps viewer) as a "standard" for stability testing when overclocking, is that everyone can get a different result depending on HOW the test is run. Let me explain. This app, in the hands of the many, is prone to giving false positives. The problem is not the app but the execution of the test itself. People with safe overclocks on a 4870 crashing in under 2 seconds while all other test are ok end up falsely assuming that thier overclock is NOT ok and some mysterious part deep in the bowels of the card must be overheating. While on the other hand, someone thinks because they can loop Furmark for hours falsely assume they are ok when in fact the overclock could be a little high for the long run.
How do these false positives occure? Easy. Here are several examples of how to spoof the app and the user...
EXAMPLE 1 - Apply 4xAA to a fullscreen or windowed test and the framerate drops. Framerate drop = less load on CPU even though it says it's still at 100% utilization of the GPU. Lowered GPU load in this case = lower GPU temp and as a result the VRM's and other compenents are cooler as well.
EXAMPLE 2 - Running some resses are not stressing the card. On my 4870 at stock clocks and low airflow, I can run full screen without AA @ all resses indefinately no matter what temperature the room is. Except one. 1680x1050. If the room is over 24C, the test will cause the card to initiate a VPU recovery in aproximately 3 minutes and the result is a blackscreen or reboot.
EXMAPLE 3 - Running Furmark in a window as is the case with MANY overclockers as they have Rivetuner or GPUz running beside it. They merrily sit there watching the screen thinking... Wow. Am I ever stable. No. If the third party app is in the back with Furmark window "focused" or "in the front", it seems to slow the card down slightly. A few windowed tests can crash the card via HEATUP when it's running alone but open up GPUz or CPUID HW Monitor in some cases with Furmark in the front and it cools down somehow. Not sure why.
EXAMPLE 4 - The one I gave in the first post. Slight change to clocks resulting in Furmark's inability to run at all. Wit hit immediately rebooting the computer in 2 seconds or less without any heat to meantion. This is a freshly started computer with a cool card. Hit desktop and start test and it reboots.
. Not because of heat but due to a crash in the renderer. Then this perpetuates into people running all over the internet saying "YA! My 4870 just crashed in 2 seconds! The VRM's are overheating!!!"
As a result of these examples, many people with both brands of cards not pushing the card with Furmark even though they think they really are. This is not a problem in and of itself and Furmark as an APP is not to blame. The problem is that people are trusting it as a "standard" for testing overclocks. Which is potentially can be. However what IS at fault is the execution of the app as a test as there is no basic locked down control point that everyone can agree to as the "if you pass this test you are golden" standard method.
Does this make sense?
I must also conclude that my results show that "as a standard", GPU Caps Viewer run at 600x600 (default settings) is a far better test when testing overclocks effect on temps and tripping the card. But that alone will not suffice. You also need a trusted artifact scanner which can test for rendering errors as well. That little pixel screaming out that in the long run your gonna toast this card early and you better back down a notch.
I am only presenting the data as I have seen it. I have limited experience in these matters so I am not claiming that I know it all or my data is 100% fail proof. Just how I find anomolies in what people are saying about this app.