Folding Forum

My Titan Rig got 2 Core 15 P8018 4.5 Hours Run Time 3400 PPD and 5757 Credits Each Today.

Based on a GTX Titan
Core 17 P8900 4.5 Hours Run Time 191,035 PPD and 35,598 Credits Each.
Core 15 P8018 4.5 Hours Run Time 3,400 PPD and 5,757 Credits Each.
A Loss of 187,635 PPD and 29,841 Credits Each.

I am setting the client-type back to beta.
I know that FAH gets mad at us if we are NOT on the Beta Team and we use the Beta Setting for the client-type.

What else are we to do?

Core_15 has not reached its end of life yet, research projects that started up using it are still in progress and have many WU's left to complete. Core_17 coming out does not cause existing work to just end as the work needed to switch in mid-research probably can not be justified. As for getting Core_15 work, you are not the only one who has been getting mixes of WU's using both cores. There can be a number of reasons for not getting Core_17 work, they range from temporary unavailability to Core_15 having higher priority at the time you get a new assignment.

P.S. F@H does not get mad at persons using the beta flag who are not on the Beta Team, however you will not get any support here in the forum if you do. It also does slow down beta testing however as non-beta members rarely monitor their work and report problems their use of the beta flag keeps actual beta testers from receiving beta WU's to evaluate.

Then this Really Means that Core 17 are made for AMD GPU's and that is why NVIDIA GK106 and below no longer run well with NVIDIA's Current Drivers or any Driver passed 326.80.
Looking more like FAH Fixed the problem with AMD and then broke NVIDIA.
Maybe now FAH needs to be "brainstorming fixes for this issue internally on what happen to NVIDIA GPU's"

"One outstanding issue for Core17 is that while it has much greater performance for most AMD GPUs, it has been reported that PPD is slower for 5xxx class AMD GPUs. We have been brainstorming fixes for this issue internally, but we have no solution for improving this unfortunately. We are continuing to discuss this internally to see what we can do."

But now from what I have seen now for over the last year of Folding is the longer a Project takes the more cost for electric power because they run hotter and the smaller the Credit you end up with and the Cost is real high to run.

So then Short Projects run cooler take less elect power and have the higher Credits and the Cost is lower to run.

It seems that it should be the other way around.

bcavnaugh wrote:Then this Really Means that Core 17 are made for AMD GPU's and that is why NVIDIA GK106 and below no longer run well with NVIDIA's Current Drivers or any Driver passed 326.80.
Looking more like FAH Fixed the problem with AMD and then broke NVIDIA.
Maybe now FAH needs to be "brainstorming fixes for this issue internally on what happen to NVIDIA GPU's"

"One outstanding issue for Core17 is that while it has much greater performance for most AMD GPUs, it has been reported that PPD is slower for 5xxx class AMD GPUs. We have been brainstorming fixes for this issue internally, but we have no solution for improving this unfortunately. We are continuing to discuss this internally to see what we can do."

F@H didn't break anything. F@H has no control with what Nvidia or AMD does with their drivers. F@H is a passive recipient, just like you, are and have to adapt to what is given, just like you do. You are complaining to the wrong people, so try contacting Nvidia. If you expect folding to fix Nvidia's performance issues, you may end up waiting for a very long time just like AMD did.

bcavnaugh wrote:Then this Really Means that Core 17 are made for AMD GPU's and that is why NVIDIA GK106 and below no longer run well with NVIDIA's Current Drivers or any Driver passed 326.80.
Looking more like FAH Fixed the problem with AMD and then broke NVIDIA.
Maybe now FAH needs to be "brainstorming fixes for this issue internally on what happen to NVIDIA GPU's"

No, what it means is that nVidia's drivers are not optimized for OpenCL, and that in their latest releases they improved performance for more recent model GPU's while apparently worsening it for older ones. That last is only an issue if you always update to the latest drivers for whatever reason. As for F@H, they don't write the drivers, the cores have not changed in the meantime, just the OpenCL support provided by nVidia. So you can't say F@H "broke nVidia".

As for the points from Core_15, they were re-benchmarked over a year ago. But they were not made eligible for the Quick Return Bonus. Eventually the project using it will finish, and in the meantime work is being done to bring out more Core_17 projects.

bcavnaugh wrote:But now from what I have seen now for over the last year of Folding is the longer a Project takes the more cost for electric power because they run hotter and the smaller the Credit you end up with and the Cost is real high to run.

So then Short Projects run cooler take less elect power and have the higher Credits and the Cost is lower to run.

It seems that it should be the other way around.

Small projects have fewer atoms thus have fewer force calculations to do in parallel so with a higher-end video card it may have more shaders/compute units than work to fill them up with. If some shaders/compute units have nothing to do it will run cooler.

There is a pattern that I have observed over time between size of the WU and the capability of the video card. Early WU's tend to be small, and work very well in smaller video cards while being inefficient for larger ones. As time progresses, the WU's tend to get bigger, resulting in more efficient use of larger cards. At some point the WU's gets big enough that the small video cards can't do it all at one time and they become much less efficient but the bigger cards continue improving their efficiency till they peak out eventually too.

P5-133XL wrote:

bcavnaugh wrote:Then this Really Means that Core 17 are made for AMD GPU's and that is why NVIDIA GK106 and below no longer run well with NVIDIA's Current Drivers or any Driver passed 326.80.
Looking more like FAH Fixed the problem with AMD and then broke NVIDIA.
Maybe now FAH needs to be "brainstorming fixes for this issue internally on what happen to NVIDIA GPU's"

"One outstanding issue for Core17 is that while it has much greater performance for most AMD GPUs, it has been reported that PPD is slower for 5xxx class AMD GPUs. We have been brainstorming fixes for this issue internally, but we have no solution for improving this unfortunately. We are continuing to discuss this internally to see what we can do."

F@H didn't break anything. F@H has no control with what Nvidia or AMD does with their drivers. F@H is a passive recipient, just like you, are and have to adapt to what is given, just like you do. You are complaining to the wrong people, so try contacting Nvidia. If you expect folding to fix Nvidia's performance issues, you may end up waiting for a very long time just like AMD did.

I am not complaining I only made a statement about what I have seen over the last year as a new Folder.
NVIDIA is not going to bother with FAH users and I agree, they build drivers for gaming and as with current game companies they work WITH NVIDIA not blame them.

P5-133XL wrote:

bcavnaugh wrote:But now from what I have seen now for over the last year of Folding is the longer a Project takes the more cost for electric power because they run hotter and the smaller the Credit you end up with and the Cost is real high to run.

So then Short Projects run cooler take less elect power and have the higher Credits and the Cost is lower to run.

It seems that it should be the other way around.

Small projects have fewer atoms thus have fewer force calculations to do in parallel so with a higher-end video card it may have more shaders/compute units than work to fill them up with. If some shaders/compute units have nothing to do it will run cooler.

There is a pattern that I have observed over time between size of the WU and the capability of the video card. Early WU's tend to be small, and work very well in smaller video cards while being inefficient for larger ones. As time progresses, the WU's tend to get bigger, resulting in more efficient use of larger cards. At some point the WU's gets big enough that the small video cards can't do it all at one time and they become much less efficient but the bigger cards continue improving their efficiency till they peak out eventually too.

"have nothing to do it will run cooler."
It is more then or like on my GTX 690 they will run at 1200 Watts total and temps on water 50C to 55C for Core 15 and a Core 17 run at 800 Watts and temps at 38C to 42C. *This is on 3 GTX 690s so it is a total of 6 GPU running, no SMP run on this rig. Because when this rig gets 6 Core 15 it will power down so I had to add a 2nd power supply to power 2 of the 3 GTX 690 *750 PSU and one GPU and the Computer runs on a 1200 Watt PSU.

bcavnaugh wrote:...NVIDIA is not going to bother with FAH users and I agree, they build drivers for gaming and as with current game companies they work WITH NVIDIA not blame them.

Please note that PG will work with Nvidia as soon as they release CUDA JIT Compiler and if feasible, have a new FahCore for Nvidia GPUs which will significantly improve the performance of supported Nvidia GPUs.

Rumors say that CUDA JIT will be released sometime this year so let's wait and see what happens after that.

EDIT -> Clarified a statement.

PantherX wrote:

bcavnaugh wrote:...NVIDIA is not going to bother with FAH users and I agree, they build drivers for gaming and as with current game companies they work WITH NVIDIA not blame them.

Please note that we will work with Nvidia as soon as they release CUDA JIT Compiler and if feasible, have a new FahCore for Nvidia GPUs which will significantly improve the performance of supported Nvidia GPUs.

Rumors say that CUDA JIT will be released sometime this year so let's wait and see what happens after that.

Thanks, This is really nice to know about.

I have not read about the CUDA JIT update.

I did think though that CUDA was for older 5xx cards and 6xx and 7xx cards were more using OpenCL for DC.

4) CUDA is usually faster on pre-Kepler GPUs (GTX 5xx and older).
5) OpenCL is usually faster on Kepler GPUs (GTX 6xx and newer).

Properly tuned CUDA is faster than OpenCL across all NVidia cards. CUDA is NVidia's proprietary compute interface, while OpenCL is used on AMD and CPUs (and others). NVidia has full control over CUDA so they tweak every last ounce of performance out of it. OpenCL they have less control over so it is not optimized (and they probably even "de-optimize" to encourage the use of CUDA ). The reason on F@h it seems the new cards do worse on the CUDA core (Core 15) is because it is such old code and it needs to be optimized for the newer version of CUDA and newer cards (Kepler architecture).

bcavnaugh wrote:...I did think though that CUDA was for older 5xx cards and 6xx and 7xx cards were more using OpenCL for DC.

4) CUDA is usually faster on pre-Kepler GPUs (GTX 5xx and older).
5) OpenCL is usually faster on Kepler GPUs (GTX 6xx and newer).

In addition to what EXT64 has stated, currently F@H is using OpenCL for AMD/Nvidia GPUs. This makes it easy to support other OpenCL devices, in theory at least.

However, once Nvidia releases CUDA JIT, we may have a new FahCore aimed specifically for Nvidia GPUs (http://folding.stanford.edu/home/welcom ... core-17-2/). Your comparison of current FahCore_15 to FahCore_17 doesn't mean much since the CUDA implementation isn't a JIT one. Trying to update it would be a waste of resources since it would need to be "tweaked" (or even rewritten, am unsure) again once the CUDA JIT is released. Thus, it is much better if we continue to optimize OpenCL FahCore_17 and once CUDA JIT is released, possibly release a new FahCore.

To get a somewhat fair comparison between OpenCL and CUDA, you can see the results produced by FAHBench here which compares the output of OpenCL to CUDA (viewtopic.php?p=245170#p245170). Please note that the real results could be different, hopefully, in a positive manner.

With that then why are older GTX 6xx cards with GK104 and GK106 on Core 17 10,000 to 20,000 lower then the GTX 780 and above with the GK110.
I can see where the GTX 5xx cards would be lower and with Core 15 would run better, but this year NVIDIA is coming out with GTX 8xx cards and the 800m is out now.
So with what as said above should we move off of NVIDIA Cards and move over to AMD Cards to Fold?
It seems that FAH is at least three years behind NVIDIA and are now up to date with AMD GPU, not only this but a lot of Folders have moved over to the Coin Mining Bitcoin that is and on AMD GPUs

The reason that non-GK110 perform poorly with the latest WHQL release is due to the driver issue, FahCore_17 remained unchanged. Last I heard, it was being investigated (http://www.reddit.com/r/Folding/comment ... ot/cdokfz2).

If you have Fermi GPUs and want to avoid FahCore_17, using GPU Client v6 is a viable alternative.

I personally don't rely on the naming scheme that much of either AMD/Nvidia since they tend to re-brand GPUs and market it as new. Instead, I prefer to concentrate on the architecture and it seems that Maxwell will first be released as GTX 750 Ti (http://www.guru3d.com/news_story/first_ ... 50_ti.html). Let's wait and see what happens in the near future.