Wierd results with my nForce2 mobo and 5900nu.

[Malichite]

I have been experiencing some wierd results with my new nForce2 (Abit NF7) motherboard and my 5900nu. I am hoping that someone here might have a clue as to what is going on. Basically I upgraded from my old Asus A7V333-RAID (Via KT333 chipset w/4XAGP) to a new NF7 rev2.0 nForce2 Ultra400 chipset.

In general I have been pleased with the upgrade after some intense memory timing/FSB tweaking. Going through my normal game/benchmark suite I noticed that most gave 10-15% increase in performance, with one major exception. Using the HardOCP UT2k3 benchmark I am getting a major hit in the ctf-face3 benchmark (i.e. 90fps with my old motherboard to 60fps on my new one). What is very strange is that this is supposed to be the best benchmark to exercise the AGP bus since it uses very large textures (i.e. seem to recall someone saying aroung ~135M thus is *must* use the AGP bus). Thus I would have assumed that upgrading to 8X would have at least stayed the same if not increased. Figuring something was wierd I tested in AGP3.0 4X mode and AGP2.0 4X mode (taping over AGP A3 & A11 pins forces the motherboard to detect it as a AGP2.0 spec card). However, in both cases the 4X mode simply offers the expected %5 drop from the 8X mode.

Long story short, I was hoping I might be able to find someone here that could test thier 5900nu/Ultra/5950 with this benchmark and give me their results on thier AMD motherboard combination (hopefully a VIA chipset/nVidia chipset). Obviously processor and motherboard combinations with affect these, but the disparity in my test is so large that I should be able to tell if I have a unique issue. Thanks,

Since everyone asks, here is my new system spec:
AMD TB Athlon 2.2G (10x220)
Memory timings 2.5:3:3:11 (Yes, 11 is a quirk with the nforce that give better perf)
Abit NF7 nForce2 rev2.0 (AGP 8X)
GFFX 5900nu (400:850)
SB Audigy
4x7200 40G drives

BTW I did test this on my nForce at 13x166 2:2:2:5 (like my old ASUS) and it falls almost 30% lower on the ctf-face3 test. Over-clocking the NF7 lessened the gap, but I am still 25% slower than a lower clocked/reduced memory bandwidth/lower AGP spec'ed ASUS motherboard.

-Mali

[The Baron]

The 11 thing--I think it's fake. I tried it a while back when people were first going ga-ga over it, but while my Sandra scores increased noticeably, everything else dropped. Everyone on the nForcers forum just noted an increase in Sandra scores; nobody seemed to try anything else to test it. And yes, Shirley, it does make a large difference.

[Malichite]

Unfortunately I haven't played around with the 11 timing enough to see whether it impacts anything more than Sandra/MemTestx86. I finally settled on 11 since the artificial benchmarks shows a slightly better score, but also seemed to run more stable. That being said, I have run some more tests with UT2K3 and wanted to get other opinions.

Running the ctf-face3 without AA/Aniso the NF7 bests the Asus A7V333 by the expected 10-15% shown in all the other tests. However, once I use fairly heavy AA/Aniso (at least 4XAA or higher and 4X Anisotropic or higher) then the numbers begin to cramp. Since extra texture size is required to do this it appears above some threshold it handles the increased size less efficiently than the KT333 chipset. The question that is unclear to me it whether this is a AGP bus issue or perhap the memory controller. For curiosity sake, I am going to test this again using single channel memory.

Baron, do you have UT2k3 so you could give me your ctf-face3 score at 4xS/4xAA. With all my tweaking I have been able to pull close to 73fps, while with my ASUS I was near 100fps at a lower clock rate/less memory bandwidth/slower AGP. Since I am mainly concerned with the PCI lock/higher multiplier, I am trying to decide if I should try a KT600 chipset instead.

-Mali

[Cotita]

I also have an abit nf7-s with xp2500+,a bfg asylum 5900nu and 2x512 mb ram. I'm building it today.

Is your problem with ut2k3 only or do you have issues with other games?

BTW from my initial testing 2.5-3-3-11 performs as fast or even faster than 2.5-3-3-5 plus it overclocks better and is more stable. Of course, your milleage may vary.

[Malichite]

Yup, after serious FSB/mem tweaking, currently the only issue is limited to ctf-face3 map with AA/Aniso, but I don't believe the issue is limited to UT2K3. I believe the issue is related to high volume of traffic across the AGP bus. Is anyone else aware of another program that stresses the AGP bus? I can't even find a free synthetic benchmark that is up to date enough to stress the AGP4/8X bus.

BTW I tested it again using my memory in single channel mode, but simply got the expected 5% drop in performance.

-Mali

[Malichite]

Well it appears the wierd quirk was caused by the position of the memory DIMMs. The Abit Manual specifies to place DIMM in slot 1 & 2 to have single channel or 1&3/2&3 for dual channel (they also included a subtext that said they recommend filling DIMMs slots in a 3 - 1 order). Either way I was using slots 2&3 and as far as CPU-Z, WCPUID, MemCache, MemTestx86, and Sandra were concerned I was in dual channel mode and got dual channel bandwidth numbers. As a last try I swapped the memory in DIMM slot 1&3 and although I recieved *exactly* the same data from the programs above, I now recieved my normal GFFX performance (actually 4-5fps faster with 4XS/4xAniso) on UT2K3 ctf-face3 map.

Either way long story short, it appears to have been the memory all along and nothing related to APG3.0. Had replies from other forums that tried to blame it on the missing SBA option or an actual problem with my GFFX, but the solution was simpler. I tried different memory position when I was first installing, but since 1&3/2&3 gave identical results I decided to follow Abit's recommendations and that was the cause of all my headaches with the large texture traffic caused by that UT2K3 map. Thought I ought to reply in case anyone else experiences the same problem.

-Mali