Folding / Tweaking a 1660 Super

[Endgame124]

Since I haven’t wanted to abort processing on my 1660 super for testing with the bench software, I adjusted the memory speed while processing a real WU. It’s not very scientific, but its interesting.

At stock, my power target was near 100% - memory at 6801 MHz and gpu clock at 1850, and I was processing at 679000 points per day (with a small variation).

Decreasing memory -250mhz briefly decreased board power, but the gpu quickly jumped to 1900mhz keeping board power near 100%. This increased estimated ppd to around 690,000.

Decreasing memory to -500mhz decreased board power again and the gpu went up another 50mhz (with small variation). Estimated ppd went up to 700000 with small variation.

Decreasing memory to -750mhz has the gpu occasionally jumping to 2000mhz, but it doesn’t stay there. Estimated ppd went back to 690k.

Decreasing memory to -1000mhz has no effect on gpu speed and estimated ppd went down to 660000. -1000 was as low as I could take the memory with EVGA precision X, so I might need to see if other software lets me go lower. I would like to see ppd at something like 4750.

Again, not scientific and only 2 minutes of observations on a live WU which will have variation on its own, but there appears to be a benefit to trading memory speed for gpu speed, to a certain point.

does anyone know if the other over clocking software (ASUS, Msi, etc) will allow a memory down clock less than -1000mhz? I would kind of like to try -5000 if it’s possible.

 

[TennesseeTony]

AfterBurner only goes down 500.

 

[Endgame124]

AfterBurner only goes down 500.

Good to know! I’ll try gpu tweak next then!

 

[Endgame124]

Of the 3 programs, EVGA Precision 1 gives the most control - gpu tweak allowed -800.

Playing with the newest version of FAH bench yielded this:

WU dhfr
atoms 23558
5 minute run time

Stock Memory (6800) / stock GPU (boost 1905)
Score: 77.936

+600 Memory (7400) / stock (boost 1905)
Score: 79.543

Stock Memory (6800) / +170 GPU (boost 2070)
79.6323

-1000 Memory (5800) / +170 GPU (boost 2070)
score 79.49

+600 Memory / +100 GPU (boost 1995)
Score: 80.73

WU real
atoms 64614
1 min

Stock Memory (6800) / stock (boost 1905)
score: 51.239
Scaled: 154.621

Stock Memory (6800) / +170 GPU (boost 2070)
Fail

+600 Memory (7400) / stock (boost 1905)
score: 51.976
Scaled: 156.846

-1000 Memory (5800) / +170 GPU (boost 2070)
score: 52.544
Scaled: 158.561

+600 Memory (7400) / +100 GPU (boost 1995)
score: 53.519
Scaled: 161.502

 

[Assimilator1]

Interesting, I found in the past that MW@H hardly cared at all about memory bandwidth, on the HD 5850 I had I would drop RAM clock to the Min (by 1/2?) & it would have no effect on times & give a nice drop in power usage . Didn't know that grx RAM could be dropped at least a little with minimal impact on F@H.

 

[Endgame124]

Interesting, I found in the past that MW@H hardly cared at all about memory bandwidth, on the HD 5850 I had I would drop RAM clock to the Min (by 1/2?) & it would have no effect on times & give a nice drop in power usage . Didn't know that grx RAM could be dropped at least a little with minimal impact on F@H.

I think the GFX ram speed changes are going to depend greatly on the GPU being used. The baseline GTX 1660 is memory starved, so I wouldn't decrease memory speed on one at all, though the 1660 super may have a little more bandwidth than it really needs (it has 10% more bandwidth than the 1660 ti despite having 10% fewer cuda cores). The 2060 Super is likely in the same situation as the 1660 super - you could probably reduce the memory speed with minimal impact. The only power meter I have is the one on my UPS the host is connected to, so I don't have a perfect way to measure power usage of the 1660 super, but I suspect that dropping the memory clock likely gives it a very solid points per watt rating.