Need help tuning an Asus 3090

[Charlie5277]

Hi, folks - just added an Asus ROG STRIX 3090 OC GAMING card, and I'm a bit disappointed in the performance. Here's a screenshot from my cell phone:

The temperature is considerably elevated, despite my increasing the fan speed from 90% to 100%. And the speed is down around 111; should be up around 117. I'm using MSI afterburner for overclocking, and the OS is Win 10. I have the core clock set to -150, the memory clock to +900, the power limit at 83 and the fan speed manually set at 90% for the other 2 cards, 100% for the Asus. I'm also getting random system crashes, about 1 per day. System was rock solid prior to adding the Asus card.

Anyone have any suggestions? >Charlie

 

[Nebell]

Try running it solo and see if you're getting the same issues?
No idea what your build looks like. But it could be a number of things. Bad bin, PSU not enough, 3rd card adding extra heat...

 

[Charlie5277]

Thanks for the fast reply, Nebell. Build is a standard open metal frame with 3 - 3090s spaced about 6 inches apart. I could increase the spacing, see if that helps the temps. PSU is a 750 watt unit dedicated to just the Asus 3090 - the MB is powered separately. When you say bad bin, you mean just a bad batch of cards?

 

[Schro]

Try testing a little lower on the memory speed and see if that improves your hash rate. GDDR6X is ECC memory, so in many cases, if you go slightly too far, you'll just get slower performance with the occasional crash (as it's correcting itself due to the ECC) instead of a quick crash into a wall that you get with non-ECC memory.

I have that card running in my desk doing about 116MH/s using 80% power, -400 core +800 memory (in afterburner). Fans are on auto, but running at 85%, maintaining 54-58c.

The other thing I've found with 3090s is that power limit can get a bit weird with some cards. For example, 90% may be just fine, but 85% drops 10-15% off the hash rate, even with the same core settings. First see if you can stabilize at the desired hash rate and stability at stock, then change one variable at a time....

 

[motqalden]

I have this card and the memory is in throttle territory even at 100% fan speed . Back side gets scorching hot and it ran much better when i pointed a fan at its back side. Edit sorry i have the Tuf version

 

[applegrcoug]

Those 3080s and 3090s with the ddr6x just get hot. I have two tuf 3090s.

One is in my regular PC and seems to get 110-115mh in windows. I set the OCs in the miner. I also put another heatsink and fan on the back and my ram still is in the mid 90s. Without it, it would throttle immediately.

The second card is in a bykski full cover waterblock running hiveos. Unfortunately hiveos doesn't display ram temps, but I know it is cool...downright icey now. I get 126mh. I also have a 3080 that does 108mh.

I have another bykski for it as well, just haven't installed it yet.

 

[She loved E]

OP did you get your 3090 sorted? Curious if you were able to tune core & mem settings for better performance.

 

[Charlie5277]

Hey, everyone - wanted to give you all a big thanks, and a status update. Sorry for the lengthy delay; had a lot of drama in my personal life - not the good kind - that I had to deal with.

Schro suggested that lowering the clock settings might improve performance, due to the DDR6 memory being ECC. Schro is correct; there's kind of a bell curve for performance, and increasing mem speed can result in worse performance, without crashing the card. But as it turned out, through sheer blind luck I had my mem speed optimized.

motqalden suggested pointing a fan at the back side of the card. This prompted me to download hardwareinfo and check the memory junction temps. I was horrified to discover that the memory temps were running at 105 - 108 C! The other cards weren't much better, at 98 - 103 C. I immediately ordered 3 Noctua fans, connected them to the MB, went into the Bios and set the fans to run at 100%. I then spaced all the GPU cards out a lot, and aimed the fans directly at the back of each card. This helped the temp situation a lot, and probably greatly extended the life of the cards. Thanks for the nudge, motqalden. You rock.

I then did a lot of experimentation with the power level and memory clock settings. What I found was that the memory clock, as Schro suggested, had to be optimized, and higher memory speeds could degrade performance. But I also found that increasing the power level would 'shift' that bell curve higher, so that the memory optimization point would move upwards. This proved to be the key factor.

Here are the final settings I arrived at:

EVGA 3090 hybrid: core clock -150, memory clock +1000, power level 84, fan speed 100%.
MSI 3090: core clock -150, mem clock +1000, power level 83, fan speed 100%.
Asus 3090: core clock -150, memory clock +1000, power level 87, fan speed 100%.

All cards are reliably delivering 118.XX megahash, giving the rig a throughput of about 355 MH. The rig is running 24/7, and is very stable; I get perhaps 1 crash per week (or less), and it's not at all clear that the crash is due to the graphics cards (I'm running win 10). Planning to shift over to HiveOS once my life calms down a little.

One note re hardwareinfo: I found that when I was running that software, the system would crash every 2 or 3 hours. Eliminating that software solved the crash problem. So use hardwareinfo to check your temps, but if your system crashes, don't immediately assume it's your overclock settings. Stop hardwareinfo, and try again with the same settings. It will probably prove to be much more reliable.

Hope that helps answer any lingering questions, and bring this thread to a satisfying conclusion. Merry Christmas, everyone. And remember, if Santa brings you a lump of coal, coal can be burnt and that will help reduce your heating costs. Peace, everyone - Charlie