Best GPU memory test?

[neurotix]

Hey all,

 

What's the best way to test OC'ed GPU memory?

 

I'm pretty confident in my limits with core clock on my 4090. However, I have the memory at 12110MHz effective which is +1650 in Afterburner. Stock is like 10500. I know it can do more as I've folded with the memory at 12210, but I'm unsure the easiest/fastest way to find the limits of my card's memory.

 

Any advice appreciated and thanks in advance.

[The Pook]

windowed Unigine Heaven. scenes still render while paused. 

 

bench until the card gets to max (VRAM) temp, pause on any static screen with space (so FPS is as consistent as possible), and watch the FPS counter while you increase frequency. Go until either it freezes/crashes/FPS drops. 

 

if the 4000 series is anything like the 3000/2000/1000 series then performance will fall off before you get visual issues/crashes. 

[neurotix]

1 hour ago, The Pook said:

windowed Unigine Heaven. scenes still render while paused. 

 

bench until the card gets to max (VRAM) temp, pause on any static screen with space (so FPS is as consistent as possible), and watch the FPS counter while you increase frequency. Go until either it freezes/crashes/FPS drops. 

 

if the 4000 series is anything like the 3000/2000/1000 series then performance will fall off before you get visual issues/crashes. 

Thanks so much. I'll try it tomorrow.

[Sir Beregond]

I've found Timespy to be pretty good at figuring out when my VRAM overclocks are unstable.

 

I've also found Unigine Superposition to be really good at figuring out clocks as its pretty consistent with scoring on repeated runs so its pretty useful for dialing in the VRAM (and core clocks) just right.

Edited April 2, 2023 by Sir Beregond

[J7SC_Orion]

3 hours ago, neurotix said:

Hey all,

 

What's the best way to test OC'ed GPU memory?

 

I'm pretty confident in my limits with core clock on my 4090. However, I have the memory at 12110MHz effective which is +1650 in Afterburner. Stock is like 10500. I know it can do more as I've folded with the memory at 12210, but I'm unsure the easiest/fastest way to find the limits of my card's memory.

 

Any advice appreciated and thanks in advance.

 

... search for ' memtest_vulkan ' on github ...I works great with GDDR6X as well and I use it on my 4090 for both ECC and non-ECC settings.  On a typical 4090, memtest_vulkan will draw around 400W + -

[SoloCamo]

Going to vouch for Timespy here.  Out of all the gpu benchs I've run, this one would always crash and need 10-20mhz lower on the memory to be stable.

[J7SC_Orion]

3DM's TimeSpy / Extreme, Port Royal and Speedway are very good tests for VRAM, though memtest_vulkan is a bit more sensitive in picking out VRAM OC that doesn't quite measure up, IMO. Below / bottom screen is the Giga vbios for the 4090 in memtest_vulkan, the Galax HoF vbios is similar.

 

Another interesting test is the ROG Furmark version (I usually set that to Vulkan instead of Open CL). The ROG Furmark also has a built-in artifact scanner !  Note of caution: ROG Furmark can REALLY test the outer limits of your card's power and temperature handling - so best to keep PL at or below 100% - at full bore, I have seen > 670 W on the GPU .  Finally, OCCT's VRAM tests is also pretty sensitive. 

 

FYI, for the OP's RTX 4090, it is worth remembering that VRAM on that card is a bit unusual - it likes to be warm...around 55 C for best oc performance...so there will be a best 'cold' OC result and a best 'hot' result for VRAM and those can diverge by as much as 100 MHz. 

 

 

 

Edited April 3, 2023 by J7SC_Orion

[Sir Beregond]

9 hours ago, J7SC_Orion said:

FYI, for the OP's RTX 4090, it is worth remembering that VRAM on that card is a bit unusual - it likes to be warm...around 55 C for best oc performance...so there will be a best 'cold' OC result and a best 'hot' result for VRAM and those can diverge by as much as 100 MHz. 

Interesting. How does it look with a water block on it? Does it stay cooler than that, or? 

[J7SC_Orion]

53 minutes ago, Sir Beregond said:

Interesting. How does it look with a water block on it? Does it stay cooler than that, or? 

 

...water block basically has GPU and VRAM temp very close to each other - not least as on this gen, VRAM is also physically much closer to the core. Add a wallop of thermal putty on the VRAM...

 

Below was at 26 C ambient but w/o water block, VRAM temps would have been warmer ('better') but hotspot would be too high. First pic = water-block, second pic = first boot after card install w/stock air cooler (obviously, different oc settings and vbios as well as lower ambient of 22 C to factor in)