Page 2 of 5
Re: GTX 1660Ti Compute Performance
Posted: Tue Feb 26, 2019 4:07 am
by Theodore
I think overclocking vram does something, but about as much as limiting power on a GPU.
On my 2060, running at stock speeds, I get 1.04M PPD.
Lowering power from stock 190W to 135W would surely affect ram and GPU speed, yet I still find my card folding at 1M PPD.
A power difference of nearly 30% giving a performance difference of only 4% .
Perhaps normal vram overclocks won't register, until nearing extreme levels?
Re: GTX 1660Ti Compute Performance
Posted: Tue Feb 26, 2019 7:14 am
by bruce
There's no doubt that overclocking the shaders makes a difference but there are reports from people who were able to overclock the shaders 2 or 3 steps (generating more heat) while underclocking VRAM enough to reduce the heat/power draw by the same amount, resulting it a GPU that's equally stable but earns more PPD.
Please do some systematic testing and report your results. Without altering shader speed or PCIe speed, alter only the VRAM speed in steps starting at whatever speed gives stable results and the decreasing it in several steps and running for long enough to get long-term averages in frame times. Then go back and pick some smaller steps between the ones you already tested (so you're no always decreasing the speed). Give us a report.
Re: GTX 1660Ti Compute Performance
Posted: Tue Feb 26, 2019 7:24 am
by Theodore
From a power perspective, it's interesting to see how the 1660 can run at 130W (probably around 120W when folding), and get ~800k PPD, while if I limit my RTX2060 to 125W, it averages around 900 and something k PPD.
It almost appears like both cards would be almost just as efficient as the other.
@Bruce,
I believe it would certainly be true on higher end cards, where cooling is a major factor (still being limited by 2 fans and nearly the same sized heat sink).
Mid sized, or budget cards usually don't run as hot, as they're equipped with nearly the same size heat sink and fans, yet produce less than half the wattage on heat.
On those cards, GPU as well as VRam can be overclocked to the maximum, without the fans even hitting 2/3rd of the speed,so I don't think ram overclock affects GPU overclock by much.
Re: GTX 1660Ti Compute Performance
Posted: Tue Feb 26, 2019 7:00 pm
by bruce
You missed my point entirely.
Overclocking VRAM does generate more heat ... but it generates only a trivially higher PPD. IF heat is a limiting factor, don't waste that heat capacity on power to VRAM, use it instead on power to the shaders. ... which DOES generate more PPD.
Re: GTX 1660Ti Compute Performance
Posted: Wed Feb 27, 2019 1:53 am
by Theodore
bruce wrote:You missed my point entirely.
Overclocking VRAM does generate more heat ... but it generates only a trivially higher PPD. IF heat is a limiting factor, don't waste that heat capacity on power to VRAM, use it instead on power to the shaders. ... which DOES generate more PPD.
Exactly!
Re: GTX 1660Ti Compute Performance
Posted: Wed Feb 27, 2019 2:57 am
by bruce
So there is likelihood is that overclocking memory can hurt your PPD. Why do you do it?
Re: GTX 1660Ti Compute Performance
Posted: Wed Feb 27, 2019 10:25 am
by Frisa
Overclocking VRAM increases memeory access error rate, memory controller would retry access this location until correct data returns. this would be be counterproductive as memory controller wasting time retrieve correct data, the final result would be slower than unoverclocked speed.
Re: GTX 1660Ti Compute Performance
Posted: Wed Feb 27, 2019 9:45 pm
by MeeLee
Frisa wrote:Overclocking VRAM increases memeory access error rate, memory controller would retry access this location until correct data returns. this would be be counterproductive as memory controller wasting time retrieve correct data, the final result would be slower than unoverclocked speed.
Not true!
Especially not on GTX cards, where the memory can be overclocked by large margins!
Anyone who is careful in overclocking will see an increase in PPD.
But you can't just overclock to whatever, you'd have to measure temperatures, and check voltages.
I have never ever seen a lower PPD as a result of overclocking memory!
Would I overclock incorrectly, and the card starts throttling, yes.
but the GTX cards weren't fully exploited when they engineered them; and memory overclock wasn't well researched when this newer Pascal architecture came out.
instead now, the RTX cards have much higher memory clocks from factory, and overclocking is much harder.
Partly due to DDR6, but also partly because memory overclocking has been taken into account.
These GTX cards could have easily come out of the factory with a 300-500Mhz overclock on the memory.
in any case, I still stand by the fact that memory overclocking does affect PPD in a positive way; and much more than a few PPD per 100k.
At least, in my case it does. GTX 1050/1060/1030 all benefit greatly from memory overclocking.
GTX 1030 bumps the PPD from 40k to 57k PPD with a 600-700Mhz VRAM overclock.
As far as shader overclocking, most OC tools don't offer this option, or it's either built into the memory or GPU overclocking; so I don't have access to it from my control panel.
Re: GTX 1660Ti Compute Performance
Posted: Thu Feb 28, 2019 4:15 am
by bruce
MeeLee wrote:Anyone who is careful in overclocking will see an increase in PPD.
...
I have never ever seen a lower PPD as a result of overclocking memory!
I don't think anybody is going to argue very much. What you say is true. It's possible to gain a percent or two by overclocking VRAM, but as I said above, heat/power also increase and you can gain more PPD by using that same heat margin to overclock the shaders. FAH performance depends mostly on shader speed and only slightly on memory speed.
Re: GTX 1660Ti Compute Performance
Posted: Thu Feb 28, 2019 9:32 pm
by toTOW
Our Geforces don't have ECC ... if it's overclocked too much and corrupts data, it will result in a bad state detected on GPU, a rollback to previous checkpoint will occur and it will result in a huge loss in efficiency. Since the margin in overclockng the memory is tight before starting to get errors, the gain of a little PPD increase from memory OC is not worth the risk.
Re: GTX 1660Ti Compute Performance
Posted: Fri Mar 01, 2019 12:14 pm
by Theodore
We yet have to see the first person posting actual PPDs on the 1660.
I'm eager to see if it's closer to a 1070, or a 2060...
Someone buy one please!
Re: GTX 1660Ti Compute Performance
Posted: Mon Mar 04, 2019 6:53 pm
by HaloJones
Theodore wrote:We yet have to see the first person posting actual PPDs on the 1660.
I'm eager to see if it's closer to a 1070, or a 2060...
Someone buy one please!
Still waiting! Someone must have one already?
Re: GTX 1660Ti Compute Performance
Posted: Tue Mar 05, 2019 8:12 pm
by toTOW
None of the new TU116 variants were supported by FAH. I just added them.
Re: GTX 1660Ti Compute Performance
Posted: Wed Mar 06, 2019 12:16 am
by MeeLee
toTOW wrote:None of the new TU116 variants were supported by FAH. I just added them.
How do you add them, without the card? Or did you get yourself one?
Re: GTX 1660Ti Compute Performance
Posted: Wed Mar 06, 2019 1:33 am
by jjbduke2004
HaloJones wrote:Theodore wrote:We yet have to see the first person posting actual PPDs on the 1660.
I'm eager to see if it's closer to a 1070, or a 2060...
Someone buy one please!
Still waiting! Someone must have one already?
I haven't folded in a while, but I have one and wanted to put it through its paces. I pulled in the latest GPUs.txt and now it's working. I'll report in tomorrow.