GTX 285 Matrix, LN2 and 3DMark 03

[bassplayer]

I figured that I would start this thread for a few reasons: first, the result is decent and second, it was damned hard to get...

I recently purchased a 285 Matrix because I haven't really played around with the Nvidia 200 series much because I wasn't into subzero stuff when it was around (and I only had AMD/ATI at the time). So I got the card all prepped with liquid electrical tape, sourced the GTX 260 Lightning / 285 Matrix voltage tool (available here at Madshrimps).

I read a few results on HWBot looking for info on the card as well as reading Pieter's article. I dropped down to -100°C and set the card to 900 core. Throughout the bench I dropped the card down to find a CB/CBB. There didn't seem to be any cold bug, but once the benchmark finished and I was brought back to the desktop, it was artifacting like crazy at -190°C until the screen froze.

Since that didn't work out, I decided to try something more a bit more conservative. I set 1000 core with 1.40V. This passed without any issue right around -100°C. Next I raised core speed to 1025 with 1.45V and and lowered the temperature to -120°C. It made it through GT1 and during GT2 when the guys are on the elevator it froze for a second. After that, the colors were odd and the score was terrible. I tried a few more things but the one that ended up working was not letting the card get colder than -110°C. This passed flawlessly at 1025 core.

I tried a few other speed/voltage combinations. 1150 at 1.525V made it all the way to GT3 at -110°C. If I let it get colder, it would fail during GT2. If it made it through GT2, it would freeze in GT3 when she's walking in the room at the beginning (one time it froze right at the end of the bench).




So the point of all of this is that I still have no idea how to bench this card. It seems like there is some sort of cold slow that moves around (no idea how or why). I tried more voltage and less voltage and all kinds of different clocks but it pretty much seemed like luck when I got a good result.



Have any of you guys benched anything in the 200 series? Does this sound like anything you recognize? Mind helping a bit?





Here's the good stuff (click for larger):

[massman]

Sounds a bit like my card too. Sometimes I can not run it very cold, sometimes it just works at high clocks.

In the last session, I think we did a couple of runs at -90°C and then slowly went down to -120°C for higher clocks.

[bassplayer]

Did you have the same types of failure when it got too cold: a short freeze followed by a bad score and some odd colors?

[Gunslinger]

Sounds almost exactly like my card as well, I think the highest I had it was 1100MHz or so for '01 and at -120 but it was not repeatable.

The color issue was probably the memory freaking out from the cold.

[bassplayer]

You think we could pull something like the 480 and hit it with a hairdryer? Or maybe kick up the mem volts to raise the temperatures a bit? I might give that a go... good idea Gunslinger!

[Gunslinger]

I tried a vmem mod on mine and it did not help it with cold.

[bassplayer]

I tried a vmem mod on mine and it did not help it with cold.

I think I'll try the hairdryer thing. With the memory on the front, will using a hair dryer on the back help keep the memory warm or do I need to find a way to apply directly to the chips?

[massman]

When we (Madshrimps team) did the 100K+ 3DM03 run, we had a lot of issues with cold too. I see I noted -100°C in the HWBOT submission, but I remember running and re-running plenty of times at -85°C before that.

During the Campus Party 2012, Der8auer had his VGA pretested to 1200+ core, but very similar issues when doing the full runs with high CPU clock.

[bassplayer]

Hitting this again today. I found out my slacker approach to LOD actually hurt my scores (I'll provide graphs and data after the competition).

I also tested my memory up to 1450 on stock volts... didn't want to go higher even with more volts.

Additionally, I tried unlinking the shaders and clocking those higher and it outright destroyed the scores. (2-3K lower even with higher shader clocks). Once I get a run I'm happy with and find the maximum frequency for my core, I'll try higher shader and see if it helps at all.

EDIT: After some reading around, people recommend unlinking shaders and clocking the core higher as the shaders impede the core OC. I'll try this as well.