RTX 4090 FE unstable because of lower VRAM temps?

[Slipknot79]

Yo,
i have made a move to watercooling today. Before that move, the VRAM has run at +1800MHz.
With watercooling even +1700MHz are not stable, games crash or i get bleck screens (PC is not frozen)

Can lower VRAM temps cause more problems with overclocked VRAMs? Is anything known about this?

 

[erocker]

It's more likely that some component isn't making contact with the waterblock.

 

[GerKNG]

1700?!
None of mine is able to run 100% stable and error free above 950.
(Single texture glitches after hours of playing a game for example)
Try port royal in 4k without ECC and you might see how "stable" it actually is

 

[3x0]

Can lower VRAM temps cause more problems with overclocked VRAMs? Is anything known about this?

Several other users have noticed the same. Some have even intentionally used bad thermal pads for VRAM just to increase their temps

 

[Slipknot79]

I have looped royal 20 times with +1800MHz with stock air cooling. (without ECC)
With watercooling i am getting instant crashes or black screens.

Another cause could be the vertical GPU mount system with a risercable? Can a risercable cause problems with VRAM stability? (Before that I have used the same mount with my 2080Ti without any issues.)

 

[3x0]

Another cause could be the vertical GPU mount system with a risercable? Can a risercable cause problems with VRAM stability? (Before that I have used the same mount with my 2080Ti without any issues.)

No, it's purely the low temps.

 

[Slipknot79]

Are there any suggestions to what pads can be used for higher temps?

 

[3x0]

A brave thing to try is to leave the plastic film on one side of the pads, that's what others suggested but I don't think anyone was (rightfully) dumb enough to try it and risk their card

 

[ir_cow]

Like someone else said, if it worked before and now it doesn't you did something wrong. Although, considering 1313 MHz is stock and 15000 MHz a decent overclock, 1800 is insane. It's hard to imagine that was ever stable to begin with.

 

[Slipknot79]

In compensation of the good VRAMs i have a bad GPU. More than +120MHz is not stable, thats around 2860MHz with air cooling.

I dont think, the risk is very high. The VRAMs can afaik do around 100°C. Now i have 36-40°C ingame. I cant imagine leaving the thin plastic film will rise the temps that much, there is still a water cooler on the other side. A nice goal could be 60-70°C.

 

[dgianstefani]

Are there any suggestions to what pads can be used for higher temps?

Just use less coverage, cut a strip half the size of each chip, the chip will still cool, but it will have half the thermal transfer area.

I also think your overclock is too high.

 

[Dr. Dro]

Due to how memory works (to simplify, it's essentially a very large array of capacitors that flip 0 and 1, it's more complex than this but this is the basics), you actually want it to be slightly warm. Not much, but just around what you'd call ambient. A similar concept applies to NAND chips used in SSDs, you want the NAND itself to be a little warm while keeping the controller cool to minimize strain on the memory cells and maximize performance.

However, +1700 seems an unreasonably excessive overclock that is probably nowhere close to stable, even for an RTX 3090 Ti or 4090 card with NVIDIA's new G6X implementation.

 

[3x0]

I've seen people stable at +1900MHz in Benchmarks and vulkan_memtest which is very good at finding VRAM instabilities https://github.com/GpuZelenograd/memtest_vulkan
@ir_cow Here's a few people from another forum experiencing the same issue, so it's not OP that did something wrong

[Official] NVIDIA RTX 4090 Owner's Club

www.overclock.net

This is funny.

+1500 Memory at 25-28C, can only watch YouTube for 10-15 minutes before green screen. But can game, stress test, etc for 4-8 hours straight at 45-50C.

Had to dial memory back to +1350 to watch YouTube for 30 minutes, no green screens. What a generation, water-cooling hurting performances.

[Official] NVIDIA RTX 4090 Owner's Club

www.overclock.net

Nope, this is the cold memory bug on the 4090. His memory cannot run +1500 cold. He should try maybe +1400 or even +1200 when not gaming and memory is cold. On my 4090, I can bench and play at +1850 all day. But If I keep that overclock active, and the memory gets to below 30C (while working or browsing internet), my screen goes black. At +1500, doesn't matter how cold my memory gets, it remains stable an no blackscreen.

[Official] NVIDIA RTX 4090 Owner's Club

www.overclock.net

[Official] NVIDIA RTX 4090 Owner's Club

www.overclock.net

 

[Slipknot79]

Ok, i have put the thin foil back on the VRAM pads. Temperatures went up from 36-40°C to 58°C.
+1800MHz dont cause any black screens. Of course there is more testing needed and I dont know anything about long term behavior.

Thx 3x0.

 

[P4-630]

Actually quite interesting.. "Too low vram temp" , you don't read that alot..

 

[thesmokingman]

That is whacked, a cold bug in normal operational temp range!!

 

[igralec84]

My Gainward passed +1800 too, also no problems in 3Dmark Will see where exactly it doesn't pass the 5 minute test, maybe the 2h test at +1800 wouldn't pass.

 

[3x0]

My Gainward passed +1800 too, also no problems in 3Dmark Will see where exactly it doesn't pass the 5 minute test, maybe the 2h test at +1800 wouldn't pass.

In my experience, 2h was enough. I've had a few errors around 1 to 1.5h when I was ~15MHz above what were known stable 24/7 OC settings. Although I haven't tested it on 4090, so I don't know how effective it is there, I would keep the tool as and additional stability test.

 

[igralec84]

In my experience, 2h was enough. I've had a few errors around 1 to 1.5h when I was ~15MHz above what were known stable 24/7 OC settings. Although I haven't tested it on 4090, so I don't know how effective it is there, I would keep the tool as and additional stability test.

Thank you for the link, i don't really see a scenario where one would need to run a very high memory overclock (maybe eth mining haha), but +1925 didn't get any errors in 15 minutes, while +1950 and +2000 got errors after a minute. Will see if my 3Dmark scores improve as i set them with +1800 and +1925 should do it for that couple of minutes

 

[3x0]

No problem. Make sure to monitor the GB/s performance in the tool as well. If you had for ex. 990GB/s at +1800 and 950GB/s at +1900 then you know somethings wrong without it showing errors Just means ECC was able to correct them with lowered effective performance

 

[Slipknot79]

Good tests are (from my experience): speed way and Warhammer Darktide.
If you get a crash or black screen very soon, you will get it because of "low" VRAM temps. This is an indication for the cold memory bug.
Since I have put a foil on the pads (dont forget to remove any air bubbles), speed way runs without crashes.
Next test will be Warhammer Darktide.

Other tests: Just play, play play. Do whatever you would do with your PC.
If you dont play, switch your 4090 to stock clock settings. If you want to play, start your oc-profile with VRAM-oc.

 

[igralec84]

It's seems Speed Way passes on +1925 but Port Royal doesn't above +1825. Haven't tried Time Spy. There's probably some error correction going on above +1800 as there's not any points increase and looking at other scores, people can't run much more than that anyway

 

[maegnificant]

I have my card at 70% power draw, with an automatic curve generation OC for the core by afterburner and 1700mhz VRAM OC. I am slightly over the average in 3DMark Benchmarks with these settings pretty happy with the TUF.

 

[eidairaman1]

No, it's purely the low temps.

Which sounds retarded in design. I'm presuming the vram volts are tanking when the temps lower.

 

[nichetcher]

First time post here so bear with me please...


So I've tried 3 4090s in terms of OC and benching on Superposition/3Dmark-Port Royal...

MSI Trio got me the best scores and I max out at 3145Mhz/12300Mhz
Gigabyte Windforce was second at 3130/12000
Founder Edition I can't break 3000/11500 without major crashing, and I got the FE because I read that the chips are binned, but I feel like I need to either RMA or find another 4090...

How many of you are dealing with this "cold memory bug" that I keep reading about? I feel like that's what's happening on this FE, but I don't know what to do here.

Thanks in advance!