Did some research for a new card among which is the 9800+ as a candidate.
A review mentioned a "video ram" bug in the 9800 series of cards. Google confirmed this.
So I was thinking "What if this problem got introduced before?" and bingo, the german Wikipedia article on the Geforce 8 confirmed my suspicions.
Fehler im Speichermanagement Die Grafikkarten und mobilen Grafikchips der Geforce-8-Serie wurden mit Treibern ausgeliefert, die einen Fehler im Speichermanagement besaßen. Dieser von Nvidia als VRAM-Bug bezeichnete Fehler führt dazu, dass immer weiter Daten in den Grafikspeicher geladen werden, so dass dieser irgendwann überfüllt ist. Dann müssen Texturen und ähnliche Dateien ausgelagert werden, wodurch die Leistung stark abnimmt. Im Normalfall sollten vorher nicht mehr benötigte Dateien aus dem Grafikspeicher entleert werden, dies passiert bei der Geforce-8-Serie nicht.
[...]
In weiteren Tests wurde dann eindeutig festgestellt, dass die Leistungsfähigkeit stark von der Speichernutzung abhängt.
Insbesondere bei den 256-MB-Modellen der Geforce 8600 GT und GTS wurden noch größere Unterschiede festgestellt, durch theoretische Überprüfungen ließ sich jedoch feststellen, dass jede Grafikkarte der Geforce-8-Serie diese Fehler besitzt.
Nvidia reagierte und versprach einen Treiber, der das Problem lösen sollte. Dieser verzögerte sich jedoch mit der Begründung, dass ein „Problem komplexer Natur“ vorliege. Ende August 2007 stellte Nvidia dann ohne weitere Ankündigung den Beta-Treiber ForceWare 163.44 bereit, dem viele Internet-Magazine unterstellten, er löse das VRAM-Problem. Dies geschah meist auf Grundlage eigener Tests, nachdem die Geforce 8800 GTS mit 320 MB in getesteten Spielen deutlich schneller wurde, die Geforce 8800 GTS mit 640 MB allerdings keine Unterschiede zeigte. Wie jedoch Nvidia auf Nachfrage mitteilte, sei in diesem Treiber der Fehler nicht behoben, die Leistungssteigerungen seien auf Änderungen zurückzuführen, die den VRAM-Bug etwas abschwächten. So sei die Speicherkompression überarbeitet worden, wodurch beim Speicherüberlauf nicht mehr so viel Leistung verloren gehe. Außerdem würden in Stalker die Texturen anders verwaltet, was dortige Leistungssteigerungen erkläre. Laut dem Technical Marketing Analyst Nvidias, James Wang, wird der nächste offizielle Treiber die Fehler im Speichermanagement beheben. Mittlerweile sind zwei weitere offizielle Treiberversionen erschienen, der VRAM-Bug ist aber noch immer nicht behoben.
Der G92-Grafikprozessor, der auf der Geforce 8800 GT und auf der im Dezember 2007 vorgestellten Geforce 8800 GTS zum Einsatz kommt, stellt eine Weiterentwicklung der G8x-Chipserie dar. Da er gegenüber dieser diverse Veränderungen erfahren hat, ist es möglich, dass der VRAM-Bug ebenfalls behoben wurde, Belege hierfür stehen bisher aber noch aus.
Translation (quick and dirty):
Errors in memory management
The video cards und mobile video chips of the Geforce 8 series were shipped with drivers, which contained an error in memory management. This error, called "VRam bug" by NVidia leads to continuous loading of data into the video memory, so that it overflows at some point. Then textures and other data will have to be put elsewhere, which leads to a severe loss in performance. Normally, data which is not needed anymore should be removed from the video memory, this doesn't happen with the Geforce 8 series though.
[...]
Further tests showed that performance is very dependant on the use of memory. Bigger differences were recorded especially with the 256Mbyte versions of the Geforce 8600GT and GTS, theoretical checks showed that every video card of the Geforce 8 series contains these errors.
NVidia reacted and promised a driver which was to take care of the problem. Yet, this driver was delayed, since this problem was "of a bigger nature". At the end of August 2007, NVidia, without announcement, released the beta 163.44 driver, which, according to many internet magazines resolved the VRam problem. This was often based on proprietary tests, in which the Geforce 8800 GTS with 320 Mb showed a significant improvements in games, but the 8800 GTS with 640 Mb didn't show any improvement at all.
On request, NVidia responded that the bug hadn't been resolved in this driver, the improvements in performance could be traced back to changes which slightly toned down the VRam bug. The memory compression had been improved, so that the preformance didn't suffer as much as before during memory overflows. Additionally, texture management in Stalker had been changed, which explained the greater performance in that game.
According to NVidia's Technical Marketing Analyst James Wang, the next official driver will resolve the error in memory management. However, at this point, two further official driver versions were released that don't resolve the VRam bug.
The G92 GPU, used in the 8800GT and on the 8800GTS, which was presented in December 2007, is a derivative of the series of G8x chips.
Since it received serveral changes compared to its predecessor, it is possible that the VRam bug got resolved as well. Proof for this is yet to come though.
Alright, this article is a bit dated (driver versions, etc...) but it has very valid points.
I don't know to what extent the G92 cards suffer from this bug (if at all), but I think I have an idea now about what goes on.
Basic contributors:
First of all, FSX as one hell of a benchmark for my 8800. Autogen maxed out, AI traffic, weather, custom planes, custom scenery. All combined with anti-aliasing and anisotropic filtering. This taxes the video memory pretty well.
Second, the hardware, an 8800 GTS 640 MB Rev 1 from 2006, featuring the aforementioned Vram bug.
Enter: The drivers. "Bug fixed" from versions 9x.xx till 17x.xx. Now there's the 180 series drivers. Optimized for NVidia's new hardware built with chips which got rid of or at least weakened the VRam bug. I suspect NVidia somehow left the bugfix out intentionally to get more performance out of their new cards (understandable, since the market share of the old G80 series chips drops over time and thus, the bug becomes less important).
Now, let's combine all of this.
I start FSX, I set up a flight, I go flying. The video memory continuously fills up, performance is normal. Then, I suddenly get an awful performance drop as the video memory overflows and FSX's game engine starts putting graphical stuff like textures elsewhere (like the Ram). Performance goes up again a bit (maybe one leftover routine in the drivers flushes a bit of old data from the VRam or FSX switches to "get stuff from the Ram while the VRam flushes" mode), before the autogen starts behaving strangely (video driver getting confused?). FSX can't receive any good data anymore and shuts down (CTD). End of story.
Other - in particular more recent - games like Far Cry 2 or Fallout 3 don't show this kind of behaviour because they either have better engines, which are able to cope with VRam overflows or don't use as much VRam as FSX (although I think I once had a similar problem with IL-2) as well as simply receiving optimizations in terms of memory management in the current drivers.
Yet, these are just assumptions based on personal experiences. The root for the problem could easily be somewhere else.
I'm really glad though that I could find a highly plausible cause for my problem.
Seems a new video card really *is* the way out of this.
Sherlock Björn, out.
