Folding Forum

iceman1992 wrote:

Endgame124 wrote:

bruce wrote:And what about the GTX 1600 series. FAH gains nothing from the extra features in the RTX series ... so unless your into AI or ray-tracing....

I have a 1660 super - completely stock in a well ventilated case, it’s pulling 700k - 800k ppd in Windows. I highly suspect that I’ll be able to dramatically down clock the ram to save power and increase points per watt. I would really like to get a comparison with someone running a regular 1660 to see if they get points in the same range. If so, we can validate that the much lower memory bandwidth of the normal 1660 doesn’t make a difference in ppd

I have a regular GTX 1660. Getting 500-630k PPD (More frequently hangs around 540k). I'm not sure if that's actually low or expected for this card. Anyone else with a 1660?

Since the GPU configuration is the same, if the 1660 gets 20% less ppd, then we would know the extra memory bandwidth is important and needs to be factored into expected points per day per watt (which is not reflected in synthetic TFlops performance expectations)

Endgame124 wrote:Since the GPU configuration is the same, if the 1660 gets 20% less ppd, then we would know the extra memory bandwidth is important and needs to be factored into expected points per day per watt (which is not reflected in synthetic TFlops performance expectations)

Yeah, but we need to confirm with another 1660 because I'm afraid my setup is somehow underperforming

Endgame124 wrote:Since the GPU configuration is the same, if the 1660 gets 20% less ppd, then we would know the extra memory bandwidth is important and needs to be factored into expected points per day per watt (which is not reflected in synthetic TFlops performance expectations)

The 1660 Super has 75% more memory bandwidth than the 1660. You could test your theory by dropping your memory clock down from 1750MHz to 1000MHz and, assuming it's stable at that rate, run a before/after test with FAHBench.

Given we have so many folders, isn't the best way to sort this to just have a new thread asking for GPU model, core clock, OS and a PPD figure. Yes they will vary according to PRCG but with enough datapoints you'll still get a good enough average.

HaloJones wrote:Given we have so many folders, isn't the best way to sort this to just have a new thread asking for GPU model, core clock, OS and a PPD figure. Yes they will vary according to PRCG but with enough datapoints you'll still get a good enough average.

I was thinking of setting up a website to collect PPD data, submitted by users, any suggestions on how I can verify that the numbers are real?

NoMoreQuarantine wrote:

Endgame124 wrote:Since the GPU configuration is the same, if the 1660 gets 20% less ppd, then we would know the extra memory bandwidth is important and needs to be factored into expected points per day per watt (which is not reflected in synthetic TFlops performance expectations)

The 1660 Super has 75% more memory bandwidth than the 1660. You could test your theory by dropping your memory clock down from 1750MHz to 1000MHz and, assuming it's stable at that rate, run a before/after test with FAHBench.

I will definitely do this when I run out of work!

It also looks like we can parse the F@H logs to mine data better than just checking in on the client every now and then as well. Need to see if someone has already done this, or if I should write my own parser.

iceman1992 wrote:

HaloJones wrote:Given we have so many folders, isn't the best way to sort this to just have a new thread asking for GPU model, core clock, OS and a PPD figure. Yes they will vary according to PRCG but with enough datapoints you'll still get a good enough average.

I was thinking of setting up a website to collect PPD data, submitted by users, any suggestions on how I can verify that the numbers are real?

there is one but it's a little underutilised at present

https://www.overclock.net/forum/55-over ... abase.html

iceman1992 wrote:

HaloJones wrote:Given we have so many folders, isn't the best way to sort this to just have a new thread asking for GPU model, core clock, OS and a PPD figure. Yes they will vary according to PRCG but with enough datapoints you'll still get a good enough average.

I was thinking of setting up a website to collect PPD data, submitted by users, any suggestions on how I can verify that the numbers are real?

Ideally the benchmark tool would interrogate the GPU using a system similar to GPUZ and then automatically submit the data to an online database

iceman1992 wrote:

HaloJones wrote:Given we have so many folders, isn't the best way to sort this to just have a new thread asking for GPU model, core clock, OS and a PPD figure. Yes they will vary according to PRCG but with enough datapoints you'll still get a good enough average.

I was thinking of setting up a website to collect PPD data, submitted by users, any suggestions on how I can verify that the numbers are real?

Aside from collecting it via software that runs on their computers and submits it for them, I don't think there is a good way of verifying the numbers. I have been planning on doing something similar to this https://docs.google.com/forms/d/e/1FAIp ... w/viewform using Google Forms, but for the new version of FAHBench. I'm not sure what the best settings for FAHBench would be however, and it seems like the current version may not be loading the GPU fully.

FAHBench is not ideal IMHO because it doesn't represent the variety in the Wunits. The larger the GPU the higher the variance AFAICT. My TitanX loves big wunits that can get all its little cores busy and hates small units that leave half of it unused.

It's even more extreme for cards like 2080ti where a big unit can push the numbers over 4m ppd while a small one can halve that.

Juggy wrote:Ideally the benchmark tool would interrogate the GPU using a system similar to GPUZ and then automatically submit the data to an online database

That was one of my future feature proposals for FAHBench viewtopic.php?f=16&t=34295

HaloJones wrote:FAHBench is not ideal IMHO because it doesn't represent the variety in the Wunits. The larger the GPU the higher the variance AFAICT. My TitanX loves big wunits that can get all its little cores busy and hates small units that leave half of it unused.

It's even more extreme for cards like 2080ti where a big unit can push the numbers over 4m ppd while a small one can halve that.

Maybe we could run multiple WUs with different atom counts in FAHBench and average the results.

HaloJones wrote:...It's even more extreme for cards like 2080ti where a big unit can push the numbers over 4m ppd while a small one can halve that.

Yep, on my GTX 1080 Ti, large WUs gets me 1.5 Million to 1.7 Million but the small ones get around 600K to 800K.

I think that taking a average may not be ideal IMO, instead a range from X to Y would provide a better picture. If I want to get constant PPD, I would go for something where the difference between X and Y is less. On the other hand, if I want the highest PPD, I know what values I am looking at on a large WU and on a small WU. That may prevent posts along the line of, I spent $$$ expecting to get 1.5 Million PPD but now I am only getting 600K PPD.

iceman1992 wrote:

HaloJones wrote:Given we have so many folders, isn't the best way to sort this to just have a new thread asking for GPU model, core clock, OS and a PPD figure. Yes they will vary according to PRCG but with enough datapoints you'll still get a good enough average.

I was thinking of setting up a website to collect PPD data, submitted by users, any suggestions on how I can verify that the numbers are real?

Off the top of my hear: would probably have to verify based on normal distribution (à la Gauß), after X submissions for the same project from different people, you could spot those that are too far up/down to (auto-)remove them from the statistic, then average the rest.
Happy to contribute data.

PantherX wrote:Yep, on my GTX 1080 Ti, large WUs gets me 1.5 Million to 1.7 Million but the small ones get around 600K to 800K.

I think that taking a average may not be ideal IMO, instead a range from X to Y would provide a better picture. If I want to get constant PPD, I would go for something where the difference between X and Y is less. On the other hand, if I want the highest PPD, I know what values I am looking at on a large WU and on a small WU. That may prevent posts along the line of, I spent $$$ expecting to get 1.5 Million PPD but now I am only getting 600K PPD.

I have heard this from multiple people so far, but do not know why this happens. Why do GPUs with higher compute units perform worse on smaller proteins?

NoMoreQuarantine wrote:

PantherX wrote:Yep, on my GTX 1080 Ti, large WUs gets me 1.5 Million to 1.7 Million but the small ones get around 600K to 800K.

I think that taking a average may not be ideal IMO, instead a range from X to Y would provide a better picture. If I want to get constant PPD, I would go for something where the difference between X and Y is less. On the other hand, if I want the highest PPD, I know what values I am looking at on a large WU and on a small WU. That may prevent posts along the line of, I spent $$$ expecting to get 1.5 Million PPD but now I am only getting 600K PPD.

I have heard this from multiple people so far, but do not know why this happens. Why do GPUs with higher compute units perform worse on smaller proteins?

wasted capacity effectively. they become no more efficient than a card with half the cores and end up with only maybe 50% utilisation. give them a really big protein with >100000 atoms and the card can get up to >90% utilisation that then allows them to get comparatively low TPF, to return far more quickly than a lower spec card and get an exponentially large quick return bonus