Folding Forum

I finally got around to finishing off my review of the 1080. This time I focused on folding efficiency. Basically I just confirmed what many people on here say...efficiency can be improved by lowering the GPU power target. I wrote this article with the intent to provide the community with something referenceable and rigorous.

https://greenfoldingathome.com/2019/05/ ... ew-part-2/

Basically I used a watt meter to log the total energy used over the test period (1 week of 24/7 dedicated folding for each power setting) and then compute an average power. PPD numbers were taken from Stanford's servers.

The efficiency sweet spot I found was a power target of 60%.

Now try to combine lower power consumption with an overclock!
The GTX series cards do get better performance when you overclock both GPU and memory.
They're quite difficult to properly overclock for folding. More difficult than the RTX cards; which the RTX cards often don't have a substantial increase in performance when overclocking the memory. (like below 1% of PPD between stock and optimal values).

I even downclock gtx 1080ti mem clock because it saves power and does not hurt PPD.

I don't know about a GTX, but on an RTX underclocking the memory does nothing to save power, or increase performance.
Whatever you gain on overclocking the GPU, you lose on memory bandwidth speed, and end up with the same PPD and power draw.

I did read that memory over/underclocking on a GTX does more to performance than on an RTX card, and I notice some PPD gain setting memory speed by 1400Mhz to 15Ghz, vs stock 14Ghz on the RTX cards.

But I see no substantial, or even recordable power gain or loss, nor any temperature changes between under or overclocking memory by up to 2Ghz on an RTX card.
Perhaps there are some principal differences between the Pascal and the Turing cards.

MeeLee wrote:Now try to combine lower power consumption with an overclock!
The GTX series cards do get better performance when you overclock both GPU and memory.
They're quite difficult to properly overclock for folding. More difficult than the RTX cards; which the RTX cards often don't have a substantial increase in performance when overclocking the memory. (like below 1% of PPD between stock and optimal values).

They were running a +175MHz GPU overclock across all tests. At least the Asus Turbo model tested had the same max TDP as the Founders Edition. Typically Add In Boad manufactures increase the power limit further which while may enable you to run a game, Possibly with errors, at a slightly higher clock, won’t net you more performance folding

It sounds like ths RTX cards may have 3 clock rates: One for Memory (which you may or may not be able to set), one for the 3D calculation hardware, which is adjustable, and one for the video refresh hardware.

The adjustments on RTX cards just show up as 'memory' and 'GPU'. Not sure what the third adjustment 'video refresh hardware' is.

Gamers care about vidwo hardware. FAH doesn't.

Based on the above discussion I did some extended testing on the Pascal-based GTX 1080 to look at what happens with PPD, power, and efficiency when over and under-clocking the memory. I started at the P2 power state (folding default) with the same +175 GPU overclock and then ran tests with the memory at +500 and -500 MHz to reveal trends. I didn't go above 5000 MHz (10 GHz effective) so as to not run into memory error correction effects.

The results are here:

https://greenfoldingathome.com/2019/05/ ... ory-speed/

Basically, memory does influence points and efficiency, but you need a big swing in the settings (say 1000 MHz) to see it. People's typical overclocks of +200 MHz or so above the P0 clock rate probably won't do much, and might even hurt performance if you run into error correction. Downclocking the memory saves a bit of power, but degrades performance and efficiency slightly.

Seems like your results show 100k PPD more (or <10%) on a 1000Mhz overclock (16% overclock).
Some GTX cards do almost 1000Mhz, mine do 600-900Mhz overclock; and it does make a larger difference on slower cards (GT(x) 1030, 1050 and 1060).

Though on my RTX card, it doesn't make any noticeable difference.
I guess 1Ghz overclock on 6Ghz is a bigger difference than 1Ghz overclock on 14Gbps memory.