Overclocking GPU memory?

A forum for discussing FAH-related hardware choices and info on actual products (not speculation).

Moderator: Site Moderators

Forum rules
Please read the forum rules before posting.

Overclocking GPU memory?

Postby ProDigit » Sun Jan 06, 2019 2:26 am

Anyone has any experience with overclocking graphics cards on FAH?

My GPUs can be overclocked between 80-200Mhz safely, which theoretically could account for about 5-10% extra performance, or 5-10% PPD + bonus points in the likes of 15-20% PPDs.

But the GDDR memory can be overclocked between 600 to 1000+ Mhz.

I was wondering if overclocking GPU memory does something significant?
If not,I'd rather keep it all stock.
ProDigit
 
Posts: 242
Joined: Sun Dec 09, 2018 10:23 pm

Re: Overclocking GPU memory?

Postby gordonbb » Sun Jan 06, 2019 7:18 am

In my testing I saw only slight increases (2-4%) in FAHBench Overclocking memory.

Overclocking the Graphics (Shader) clock will give the most change but it is tricky as WUs will vary in terms of how hard they push the GPU and a offset that appears stable can start crashing when it hits a new workload.

The main issue is heat and the non linear power draw of the GPU at higher clock speeds. I do run a mild graphics Overclocking but power limit the cards to 10 to 30W under the rated power limit to keep the temperatures down and the fans running at lower speeds.

It’s a balancing act to find the right settings that yield the optimal PPD/W.
Image
User avatar
gordonbb
 
Posts: 82
Joined: Mon May 21, 2018 4:12 pm
Location: Great White North

Re: Overclocking GPU memory?

Postby ProDigit » Sun Jan 06, 2019 8:51 am

Yes. I use EVGA s tool.
Basically overclock while folding, at the cost of a few lost WUs while doing the process.
Takes about 2 hours to set up all 3 cards, then implement a safety margin.
. The HW monitor clearly shows when the CPU starts throttling, or even when CPU speeds become unstable.
On my MSI GTX 1060, I can potentially overclock the memory more than the max allowed of 1000Mhz to 5Ghz, and it still runs stable, but dialed it back by 100Mhz to 4900Mhz.

Since my cards are on risers, I can't see how much power they use, but the GTX 1060 is on the PCIE 16x slot, and even with overclocking it never surpassed 80% of pcie power consumption.

I try to keep the cards temps down a bit, and surprisingly throttling the fan up from Auto settings of 1600RPM to about 1800, lowers not only temperature from 80c to 70 something, but also lowers power draw.
My cards, both 1060 and 1050 are running 20% higher fan speed, resulting in 10 degrees C drop , and total power consumption is down by a few watts!

Overall my 1060 can do overclocking, but it seems the card already runs at pretty much the highest GPU settings. I can add 20Mhz overclock, another 50Mhz, and I can see the core clock bouncing up and down.

Same for my passive cooled 1030, they both run on nearly the max GPU speed.

The 1050 I can overclock. Not sure if it has anything to do with it being a "GTX 1050 OC" from EVGA?

But the ram on all cards can be overclocked by at least 500mhz, sometimes to 1Ghz.
ProDigit
 
Posts: 242
Joined: Sun Dec 09, 2018 10:23 pm

Re: Overclocking GPU memory?

Postby foldy » Sun Jan 06, 2019 9:57 am

VRAM OC does not help FAH only core clock OC. Another disadvantage of vram oc is it automatically increases latency. There are some non FAH algorithms were reduced vram speed -500 Mhz helps speed because that reduces latency.

For overclocking test you can use http://fahbench.github.io/
foldy
 
Posts: 1471
Joined: Sat Dec 01, 2012 3:43 pm

Re: Overclocking GPU memory?

Postby ProDigit » Sun Jan 06, 2019 2:10 pm

Latency is directly connected to RAM speed.
Latency always stays the same, even if latency numbers increase, the overall latency is exactly the same using the same ram (only when you increase latency, but leave ram speed the same does latency increase).

Eg: when you're running 1333Mhz RAM with CAS latency of 8, running it at 1600Mhz will increase CAS latency to 10. The number goes up, but the interval between RAM flushes remains exactly the same.
It doesn't affect overall latency at all. It just means that at 1333Mhz it takes 8 ticks to clear the memory, while it takes 10 at 1666Mhz.
If you do the math:
1333 / 8 = 1600 / 10 (with a minor rounding error; mainly because a CAS latency of 8 is actually 8.333...).
ProDigit
 
Posts: 242
Joined: Sun Dec 09, 2018 10:23 pm

Re: Overclocking GPU memory?

Postby toTOW » Sun Jan 06, 2019 10:00 pm

Overclocking the VRAM is very sensitive and will usually cause more troubles (computations errors -> Bad States) than the performance increase it brings. Most folders usually discourage it.
Folding@Home beta tester since 2002. Folding Forum moderator since July 2008.

FAH-Addict : latest news, tests and reviews about Folding@Home project.

Image
User avatar
toTOW
Site Moderator
 
Posts: 8712
Joined: Sun Dec 02, 2007 10:38 am
Location: Bordeaux, France

Re: Overclocking GPU memory?

Postby ProDigit » Mon Jan 07, 2019 2:53 am

I think vram overclocking is easier than GPU overclocking.
The EVGA app has a monitor that shows when throttling start.
On my GT and GTX cards, the overclock instability usually results in an immediate drop in performance, unlike the GPU which starts throttling before crashing, and depends on card heat.

I usually stay 50-100Mhz below crashing, but the GPU frequency in the app, is controlled by the input voltage, and thermally, more like finding a balance between the two to get the right overclock.
ProDigit
 
Posts: 242
Joined: Sun Dec 09, 2018 10:23 pm

Re: Overclocking GPU memory?

Postby bruce » Mon Jan 07, 2019 3:00 am

The heat generated by overclocking the shaders can be productive, particularly if you UNDERclock VRAM to reduce the power used by GPU memory. OVERclocking GPU memory is almost always unproductive -- especially if you can use that margin to overclock the shaders.
bruce
 
Posts: 22441
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Overclocking GPU memory?

Postby ProDigit » Tue Jan 08, 2019 9:58 am

Here are my findings on OC:

Totally stock system running 4 CPU cores and stock GT 1030 + GTX 1050 + 1060:
300 Watt, ~509k PPD = 1.7k PPD/W
Overclocking VRAM memory + GPU to max settings:
310 Watt, ~526k PPD = 1.7k PPD/W
Set VRAM to idle (810Mhz), and OC GPU to max:
218 Watt, ~288k PPD = 1.32k PPD/W
Set Vram to minimal speed above idle (1030 = 2505Mhz, 1050 = 2505Mhz, 1060 = 3000Mhz), and OC GPU:
288W, 507k PPD = 176k PPD/W

It's funny to note that, when overclocking VRAM + GPU, PPDs increased, it did nothing to PPD/Watt. They were identical, since Wattage rose linear with PPD count.
If Bonus points were not included, overclocking would probably be less efficient for folding than running a system stock.

Setting VRAM to idle, was even a worse idea. It seems to bottleneck the GPU.
Despite Windows indicating GPU is running at 98% load, PPDs were just over half the norm for the card.


But once I set VRAM to the minimal speed, above idle, underclocking RAM helped performance, especially on the 1060, because it dropped temperature on the card drastically, but also ran the card at a lower power envelope, saving power in the process.

The passive cooled 2GB 1030 had identical power consumption vs stock
The GTX 1050 2GB consumed 4% less power with RAM underclocked and GPU Overclocked, vs stock
The GTX 1060 3GB used 10% less power than stock.

I presume the more RAM, and the faster the RAM, the greater the energy savings are.

The 1060 always was running at it's max of 83C,
The aircooled 1030 also was temperature throttled to 63C.
So reducing RAM speed allowed me to overclock the GPUs by an additional amount.
1030 : +50Mhz
1050 : +12Mhz
1060 : +26Mhz

The power savings per PPD is negligable (12Watts at the cost of 19k PPD in my case) just isn't worth it.
Underclocking RAM to further overclock GPU is pretty useless.
Perhaps there's a balance there, where you can get max PPDs by overclocking the GPU a bit more than stock, and running the RAM a bit below stock, but it'll take time to figure all that out.
In the mean time, I would just recommend to run your cards fully stock,
Until more data comes in that proves this statement otherwise.

In the mean time, if you're hungry for PPDs, run your graphics cards (both memory and GPU at overclocking speeds, and fold with whatever CPU threads you have left.
But if you're concerned about being an efficient folder, just fold on GPU only.
ProDigit
 
Posts: 242
Joined: Sun Dec 09, 2018 10:23 pm


Return to FAH Hardware

Who is online

Users browsing this forum: No registered users and 1 guest

cron