Run benchmark tests to see if there is scalable performance gains with that memory overclock. If it drops off after say 1,500Mhz on the memory, you know it is error correcting. This will probably take some time to do.

 

Hello everyone. I am pretty new to overclocking so please bare with me. I wanted to overclock my "Gigabyte RTX 4070 Ti Super Windforce OC" and looked up a few tutorials and guides. I use MSI Afterburner and tested for stability playing Cyberpunk. I determined that everything above 200Mhz core clock leads to crashes so I set it to +200 as the optimal value. My doubts are in regards to the memory clock though. I am able to set it to +2000Mhz without crashing but somehow it seems too good to be true. People in other forums were saying that there is no way that my GPU can run stable at +2000Mhz memory clock and that there is probably some error correction going on which in turn limits my performance. How can I check if the GPU is error correcting or not? Thanks for helping me out.

View attachment 2694220

The ideal way to check at what point the memory is error-correcting is to run a synthetic bench such as 3D Mark/Unigen Heaven/Valley. Start from stock on the memory, and keep overclocking in 50-100Mhz increments. Once you start seeing stagnating/diminished results, you've hit the limit. For e.g, the memory performance threshold on my 3080 was +750 Mhz.

 

@Kookiemonster75 @HeLeX63 Thanks guys! Do I just look at the FPS in the benchmark in real time for a few minutes while adjusting the memory clock or do I actually have to let the whole benchmark finish after every increment? (I guess that would take ages)

 

It'll be hard to discern by just looking at real-time numbers since they can fluctuate with each run. You may miss out on something, and not worth the headache. Grab a beer and let the benchmark complete for accurate results. You can also increase the multiples of increments with each run, say +100 OR 150.

 

Easiest way I've always checked for memory overclocking where it starts to drop frames due to error correction is with Unigine Heaven benchmark.

Set Heaven to max settings, windowed mode, and then once the scene is loaded, set it so you can control the "player" as if it's a first-person shooter. Point upwards so you stare at the sky and don't see anything but the sky. Move around until you find the highest framerates you can. Then don't move and keep increasing your memory OC. Your framerate will keep going up. Keep increasing the memory OC until the framerate stops increasing and starts going down - that'll be the error correction kicking in. At that point, reduce the mem OC by around 2-5 clicks/steps/increments and there you go, done.

Works every time including on my previous newer cards: 4090, 3090, 3070 Ti, 2080 Ti.

For core, do all the typical benchmark stress tests like 3-5 runs of different 3D Mark GPU tests (Nomad, Speedway, Port Royal, Time Spy, Firestrike, etc.), Unigine (Heaven, Valley, Superposition), and gaming.

For mem instability, 3D Mark Speedway would crash the fastest for me with my 4090, couldn't even get halfway through 1 run. That was before I tested Nomad so I don't know if Nomad is as good or better but Speedway was definitely faster than the others (Port Royal, Time Spy, Firestrike, etc.).

Once you have your core OC dialed in, maybe re-do the Heaven method again - not sure if a higher core OC can possibly affect things. I'd just do it again to be sure (it's easy and pretty quick).

The Unigine Heaven method, I think, is definitely the fastest, easiest, and most consistent way to find the mem OC error correcting threshold.

 

+2000 actually works for me on the 4070TiS - both 3DMark Speedway and Port Royal shows scaling up with best scores at +2000, interesting that Steel Nomad has better scores at +1800 though. Suspect it might be limited by power split between vcore and memory rails on those different tests. Not enough gaming time at +2000 to confirm long term stability but the limited time I tried last night didn't lead to any crashes in BF2042 which is notorious for crashing if it doesn't like your overclock.

 

So I did the following testing with +1000, +1500, +1700, +1800, +1900 and +2000

Unigine Superposition: judging from the scores I didn't notice any error correction going on. For some reason I got the best score with +1800 (20635 points) and the second best score with +2000 (20608 points). But I think that's neglegible, right?

In-game Cyberpunk benchmark: +2000 definitely gave the best results

OCCT stability test for 30 minutes (3D Standard): no errors detected

Github Memtest Vulkan 5 minutes: no errors detected.

Games tested: Cyberpunk and FF7 Rebirth, no crashes.

Here are the scores and the data I got in Unigine Superposition:

Can I assume that this overclock is indeed stable at +2000Mhz memory clock?

 

Looks like possible error correction starting somewhere around +1650 and +1750. From those tests alone, I'd settle on +1600, max +1700. Do the sky test in my previous post in increments of +100 starting from around +1000.

 

I did something similar in games already. Faced a wall or something and kept increasing memory clock. FPS never went down until I reached max +2000Mhz. Wouldn't this be the same as doing it in Heaven?