Overclock yields performance loss... Little help, please?

[chazdraves]

Well, I'm hoping that someone can shine a little light on what's happening because I am at a loss. I'll keep a long story short.

The two parts in question are a Venice core A64 3400+ (2.2 at 200x11) and a BFG 6800GT OC (370Core 1000RAM). The RAM was kept within 14MHz of it's 400MHz rating at all times, so I don't assume that to be a factor. Also, the HTT never went below 750 (this is socket 754, so it's 800MHz default). The mobo is a Gigabyte nforce 3 job, I forget the exact model.

Here's the test scenario. I don't own a lot of games, so I used the Command & Conquer 3 Demo. I used FRAPS to record a 2 minute benchmark for the first portion of the GDI Demo level. My settings were 1440x900 with everything up except AA. This keeps my FPS below 30 at all times since the demo caps it at 30 - obviously I don't want the cap to skew results. Feel free to criticize my test, I realize it has some holes.

Here are the settings:

BFG 6800GT OC
Stock: 370/1000
OC: 425/1049

AMD 64 3400+
Stock: 2.2 @200x11
OC: 2.71 @246x11 and +10% VCore voltage (highest it goes)

Here are the results:

Default Speeds

Min: 7FPS Max: 27FPS Avg: 19.958FPS

CPU OC'd 23%:

Min: 10FPS Max: 27FPS Avg: 19.883FPS (.04% loss)

GPU OC'd 15%:

Min: 6FPS Max:25FPS Avg: 17.583FPS (14% loss)

GPU & CPU OC'd (15% & 23% respectively)

Min: 8FPS Max: 25FPS Avg: 16.675FPS

Can anyone explain what's happening here? Under load the CPU will hit 50C and the GPU will hit 68C while OC'd. I don't see any artifacts or experience any unusual stops/halts. There are no background apps running, and I used RivaTuner for the GPU OC.

Thanks very much for your time and ideas,
- Chaz

 

[Agentbolt]

Is the card PCI-E? Use CPUID, check under Mainboard for the PCI-Express link width. A lot of the time overclocking will set it to 1X. You need to change the PCI-E speed in the BIOS to somewhere between 101-120 MHZ, and check to see which frequency gets the PCI-E link width back to 16X. I had to go all the way up to 120 MHZ to fix my 8800GTS.

 

[chazdraves]

Nope. I guess I forgot to mention that it's AGP at 8X with PCI/AGP lock set at 66MHz. Thanks for the thought though.

Anybody? I feel like I've got some rare disease no one's ever treated.

Thanks all,
- Chaz

 

[crazy4life]

My friend has the same problem so you two are not alone. Im clueless on what his problem or yours is thou.

The only thing that i can think of is you might have currupted your drivers.

 

[myocardia]

This is a simple one. Your processor is faster than your video card at stock speed, that's why overclocking it gave you the exact same results (a .04% difference is considerably less than you'll get at exactly the same clock speed). Notice how your minimum FPS went up considerably, but your maximum stayed the same. That's what happens when your cpu is faster than your video card.

As far as why your scores drop with both overclocked, did you reboot after the first test? If not, your second try should always be lower than the first, even if none of your component's speeds have changed. That's just the way that Windows-based programs work, for some strange reason, even games. Try it with the SuperPi mod, if you don't believe me. The first test will always be the fastest, assuming you don't use ClockGen to raise your cpu speed, of course.

 

[chazdraves]

Very good ideas.

I've got the secret answer to the whole question, however. After getting off work last night, I had a fair share of time to do some more in-depth testing. I'll give a big hint: it has something to do with nVidia's bloody "throttling" system on the 6800GT OC. Something I hadn't heard of before. The other half is exactly as myocardia was just explaining. The CPU wasn't a factor in the testing as the GPU couldn't keep up. Thusly, the resulting .04% swing was merely the difference from one test to the next. The GPU, however, was throttling itself down to supposedly protect the card from catastrophic damage and alien attacks which caused the ridiculously low scores despite the higher clock speed.

The answer for me (and hopefully this helps your buddy, crazy4life) was to enable a couple extra options in RivaTuner which force the higher clocks speeds and then use something less GPU restricted to determine whether or not the CPU overclock was worthwhile. I don't have the exact numbers, but Super Pi reflected an equal advantage in overclock to the time decrease at 1M. Repeat tests have shown a jump from 19.8 to 21.9FPS so far.

If anyone else has similar hardware and would like updates, just post a reply.

On that note: a hearty thank you to everyone for the assistance. I'm glad it was easier than I had feared - and I'm glad there was an explanation.

Regards,
- Chaz