For context, about six months ago I figured out that +1210 MHz on the memory is my maximum actually stable overclock. At +1220 MHz, I start getting a small flickering artifact after 40+ loops of 3DMark Speed Way. At +1250, it crashes after 40+ loops (yes, I tested different benchmarks for several hours each, increasing the memory offset in +10 MHz steps).
But here’s the weird part: I don’t see any noticeable drop in benchmark scores, even when artifacts appear or the test is on the verge of crashing.
I used memtest_vulkan only as a way to benchmark the results, you can see it by the iterations. It shows improvement even up to +1300, it only throws an error at +1350, but the performance result still goes up.
I accounted for peak memory temperature too. If I understand correctly, the official max temperature for GDDR6X is around 95C, but NVIDIA seems to allow throttling only once it hits 110C. In memtest_vulkan, which absolutely maxes out the VRAM, temps are:
- 100C at +600
- 102C at >+900 MHz
…and still, the scores don’t drop.
So I end up with this bizarre situation:
Between +1200 and +1300 MHz, performance keeps improving, but the only fully stable point is +1210 MHz.
How is someone supposed to determine stability by looking at the result if it only drops at an obviously excessive value, far beyond my actually stable +1210 example?
Or did I just get a "special" sample that behaves like this?