Low PPD for GTX 1070 + GTX 1060

[lovingfriendstar]

I've just started folding on my GTX 1070 and GTX 1060 6GB on Windows 10 recently. I have been folding 24/7 for 2 days and have already completed 30 WUs after binding my passkey. But my PPD values for both cards combined are hovering around 450,000 when I should be getting at least 650,000 to 750,000 after accounting for powerlimit I've set for the cards at 70% and 80% respectively. I know the PPD varies according to the WU but the most I have ever seen is around 500,000. It is even now showing 270,000 PPD. This seems quite bad when the GTX 1070 alone should be able to pull more than that.

Why am I getting such low PPD? Can anyone please help me figure out what has been going wrong with my setup? I'm using the latest version of F@H just downloaded a few days ago.

My specs are
Ryzen R7 1700 @3.2GHz
Gigabyte GA-AB350 Gaming 3
8GB RAM
GTX 1070 and GTX 1060 6GB using 373.06 drivers
(GTX 1070 @PCI-e x16 3.0 and GTX 1060 @PCI-e x4 2.0 speeds)

[bruce]

I'm not sure I'll be able to answer the question, but facts that also might be helpful is the effective speed of the PCIe bus (Reported by GPU-Z) for each card.
(Just because your hardware specification says it supports x16 3.0 and x4 2.0 speeds doesn't prove they're actually connecting at that speed, and 4x 2.0 is pretty slow for some WUs you might be assigned.)

Also, which projects happen to have be running. Also confirm that you've completed at least 10 WUs without any changes to either the Username or the passkey across all your machines.

This may be the same issue as the ones being discussed here whether or not you're talking about the same project.

With older (slower) GPUs, any lag introduced by a saturated PCIe bus was negligible but as the speed of GPUs has increased more rapidly than the speed of the bus, it's becoming more significant. It was easy to establish a nominal PPD when the GPU's compute engine was always the part that was saturated. The bus was always fast enough to introduce no noticeable lags, but as GPU speeds as well as the amount of data in a WU have increased, it's becoming likely that the GPU can process data faster than the bus can supply it. (Then the GPU necessarily slows down waiting for data.)

This may or may not be your issue.

The actual throughput is a result of a sort of a weighted sum of data transfer delays and data computing delays. Of course there are a lot of other factors involved, (like drivers, etc.)

[Kjetil]

You run a old driver. I use 378.78 on win 10.

[Nathan_P]

I suspect the 1060 is being bottlenecked by the PCIe bus speed. If you want to see if your 1070 is running at full speed pause the 1060 and let the 1070 run for a day or so by itself. At full speed you should get around 600k PPD - given yours is power limited you should see 450-500k from that card depending on project.

[foldy]

Are you sure both cards are under load? Maybe fah tries to run both work units on the same GPU?

[lovingfriendstar]

I'm not sure I'll be able to answer the question, but facts that also might be helpful is the effective speed of the PCIe bus (Reported by GPU-Z) for each card.
(Just because your hardware specification says it supports x16 3.0 and x4 2.0 speeds doesn't prove they're actually connecting at that speed, and 4x 2.0 is pretty slow for some WUs you might be assigned.)

Also, which projects happen to have be running. Also confirm that you've completed at least 10 WUs without any changes to either the Username or the passkey across all your machines.

This may be the same issue as the ones being discussed here whether or not you're talking about the same project.

With older (slower) GPUs, any lag introduced by a saturated PCIe bus was negligible but as the speed of GPUs has increased more rapidly than the speed of the bus, it's becoming more significant. It was easy to establish a nominal PPD when the GPU's compute engine was always the part that was saturated. The bus was always fast enough to introduce no noticeable lags, but as GPU speeds as well as the amount of data in a WU have increased, it's becoming likely that the GPU can process data faster than the bus can supply it. (Then the GPU necessarily slows down waiting for data.)

This may or may not be your issue.

The actual throughput is a result of a sort of a weighted sum of data transfer delays and data computing delays. Of course there are a lot of other factors involved, (like drivers, etc.)

I have confirmed the PCI-e speeds via GPU-Z and they're correct. The cards are also Util limited or PWR limited but not bus limited so I'm not sure if that's the problem.

Also, I have used the same user name and pass key and have folded even more WUs. The units include 10496, 13200, and several others I don't quite remember.

[lovingfriendstar]

You run a old driver. I use 378.78 on win 10.

I was using 382.XX driver before and downgraded to 373.06 after seeing that it is the recommended driver for folding for now. Not sure what's the problem here though.

[lovingfriendstar]

Are you sure both cards are under load? Maybe fah tries to run both work units on the same GPU?

It is indeed running on both devices as shown by GPU-z workloads.

[lovingfriendstar]

I suspect the 1060 is being bottlenecked by the PCIe bus speed. If you want to see if your 1070 is running at full speed pause the 1060 and let the 1070 run for a day or so by itself. At full speed you should get around 600k PPD - given yours is power limited you should see 450-500k from that card depending on project.

I'll try this later. Thanks for the suggestion.

[Joe_H]

You run a old driver. I use 378.78 on win 10.

I was using 382.XX driver before and downgraded to 373.06 after seeing that it is the recommended driver for folding for now. Not sure what's the problem here though.

That is an old recommendation, look at the dates of the posts and the reasons it was recommended last year. Those issues have been dealt with.

[bruce]

As you probably know, WIndows 10 home forces updates when Microsoft decides to. ... and when they do, they kill OpenCL.
NVidia releases and certifies new WHQL drivers long before MS distributes them, so by keeping up-to-date from NV, I avoid the possibility that MS will find obsolete ones.
I'm currently running 382.xx

[lovingfriendstar]

You run a old driver. I use 378.78 on win 10.

I was using 382.XX driver before and downgraded to 373.06 after seeing that it is the recommended driver for folding for now. Not sure what's the problem here though.

That is an old recommendation, look at the dates of the posts and the reasons it was recommended last year. Those issues have been dealt with.

Which driver is the currently recommended one now? I would like to play games on my system occasionally so indeed I want to have the latest driver but I don't want to hurt my folding performance either.

[lovingfriendstar]

Alright. I've updated the drivers to the latest 384.76 and it seems like the app is reporting correct PPD. It was also getting the correct amount of points before (verified via Stanford website and folding.extremeoverclocking.com), and it is just the client was reporting incorrectly which is about half of what I should be getting.

Again... Is 384.76 any good for folding, or should I get another version? Which is the current recommended one for folding?

[bruce]

The general recommendation has alwasy been to run the latest released version from the GPU manufacturer.

Sometimes difficulties are discovered/reported and ONLY then do we recommend that you revert to an earlier version until the reported problem is resolved.

[Aurum]

My specs are: Ryzen R7 1700 @3.2GHz, Gigabyte GA-AB350 Gaming 3, 8GB RAM, 373.06 driver, and GTX 1070 @PCI-e x16 3.0 and GTX 1060 @PCI-e x4 2.0 speed.

Are you also doing CPU folding? Your CPU has 8 lanes and 2 need to be dedicated to folding GPU cards. I look at Task Manager Performance and sometimes back off and use 5 cores for CPU folding so I don't max out the CPU loads.
I assume the two cards are in slots 1 & 4 and slots 2 & 3 are empty, "The PCIEX4 slot operates at up to x4 mode when both of the PCIEX1_2 and PCIEX1_3 slots are empty."
http://www.gigabyte.us/Motherboard/GA-A ... -rev-10#sp