bruce wrote:We need to change our semi-standard answer about overclocking. Normally we say the shader clock is important and the memory clock doesn't matter much.
That's reasonable for nVidia folks but the COVID surge has brought a lot more AMD GPUs into FAH and we should think about that more carefully. AMD made a big push sometime back about the advantages of using shared RAM.
For the AMD GPU's, that is solid thinking. Having played with mine some, I would say memory speed is just as important as shader clocks. When I built this system it was for my wife's work purposes, and we needed to get it up and running 100% stable quickly. Even though I got memory rated at 3200, I didn't have much time to do any benchmarks, stability testing, playing with the BIOS settings, etc. So when I got it up and running it defaulted to DDR4-2400, and I just let it ride since I knew she wouldn't notice and if anything it would make the system more stable. Over time as she needed it less, I upped the memory clocks to 2933, ran tests, etc, then clocked to 3200 and did the same. I would say PPD return scales almost directly with memory speed increases. And the same for the GPU shader clocks, though I haven't played with them much yet.
And I do think to some extent that shared RAM is helpful. On certain WU's with low GPU utilization on the bigger cards, this APU will just reduce memory use (size) and keep the GPU side up at higher utilization. On mid to larger atom count/complexity WU's usually GPU utilization will stay up at 95-100% and the memory use allocation grows.
And I'd assume the newer APU's are even more efficient in terms of power consumption, which is surprising even on my 2400G. When I do the quick math, it's about as power efficient when GPU folding as anything up to the 1650/1660 range in terms of PPD/watt. Of course both numbers are quite a bit lower, but the efficiency isn't bad at all. And I can GPU fold with the fans virtually silent if I want, with temps rarely getting above 50C.