I found I couldn't adjust the memory speed clocks with Precision X on a new 980. Then I stumbled on the fact that the setting will not stick if you have the GPU loaded. Exit SETI and wait a few seconds for the process's to stop and then try it. Then start SETI again.

All the best,
Aaron Lephart

TechVelocity.com

---

That may be so, but it doesn't run in the state that it is in when SETI isn't running. So when you start SETI up again, you will be back to 6GHz (I think).

---

You can use the Nvidia Inspector to change the speed of P2. Close out BOINC, launch NI and select p2 for each GPU and move the speed up to 3504MHz and then click the button to the right lower corner (I don't remember right now what it's called). Do this for each of your GPUs then close out the NI and relaunch Boinc. Your GPU speed should now be set to 3.5 GHz

Jason recommend Process Lasso to help keep keep everything running smooth. I'd go with that recommendation. I just started to use it as well and it seems to keep the drivers from crashing and resetting the GPU.

Zalster

Just got an interesting explanation for why the card is not running CUDA or OpenCL at full speed (7GHz), and is being put into P2 state by the driver to run at 6GHz.

The claim is as follows: for gaming, you want the highest possible pixel calculation rate and you don't really care if there is an occasional error, as long as you don't get visible artifacts. BUT when you are doing large scientific calculations (e.g., SETI), you want to minimize the occurrence of errors, so you don't let the card run at its rated gaming speed to prevent the occasional errors that occur then. (Sort of using the gaming card as a workstation card by downclocking).

This actually makes some sense. I don't know if it is actually true, of course, but it IS interesting.

---

Just got an interesting explanation for why the card is not running CUDA or OpenCL at full speed (7GHz), and is being put into P2 state by the driver to run at 6GHz.

The claim is as follows: for gaming, you want the highest possible pixel calculation rate and you don't really care if there is an occasional error, as long as you don't get visible artifacts. BUT when you are doing large scientific calculations (e.g., SETI), you want to minimize the occurrence of errors, so you don't let the card run at its rated gaming speed to prevent the occasional errors that occur then. (Sort of using the gaming card as a workstation card by downclocking).

This actually makes some sense. I don't know if it is actually true, of course, but it IS interesting.

So the driver developers are deliberately downclocking the card for computational purposes. If you want to use a video card for computational purposes, you buy a purpose built GPGU card such as TESLA or QUADRO. If the driver developers think that GPGU users of gaming cards wouldn't check their results for errors and then continue using a card that generates errors, they must have a low opinion of the user and think they must provide a nanny service. If a cruncher sees that his cards are generating errors, he investigates and cleans the card or downclocks the card on his own until it provides valid results. It would have been nice if the driver notes had made mention of this fact when the 900 series came out. Never a problem until the 900 series came out. In my opinion, the drivers should stay out of this area and leave it up to the self-interest of the user to use the card sensibly. I don't want a nanny overseer. My $0.02

Cheers, Keith

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Well, not exactly. They seem to be saying that gaming has laxer requirements than computing, per se. You are right, however, in that it should be our option to think about/do the downclocking. Or maybe Maxwell is more prone to this kind of error(?).

If this last is true, things may get very interesting in terms of class action lawsuits, since it implies they knowingly released a (semi)defective product. If, in fact, Maxwell has this type of defect.

Maybe it's time to get a bowl of popcorn to watch the fireworks.

---

Well, not exactly. They seem to be saying that gaming has laxer requirements than computing, per se. You are right, however, in that it should be our option to think about/do the downclocking. Or maybe Maxwell is more prone to this kind of error(?).

If this last is true, things may get very interesting in terms of class action lawsuits, since it implies they knowingly released a (semi)defective product. If, in fact, Maxwell has this type of defect.

Maybe it's time to get a bowl of popcorn to watch the fireworks.

I think the fireworks and lawsuits have already started if I recall correctly with the 970 memory issue. I haven't seen any sign of weakness yet in my 970's with overabundance of compute errors. Heck, I'm even overclocking the memory from P0 stock speed by +100 MHz with no issues.

Cheers, Keith

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I decided to track this down a bit.

First I called EVGA Tech Support; the guy I spoke to (at 2nd Level) has not heard of the downclocking of CUDA apps, but he said he would look into it.

Then I called NVIDIA Tech Support, and am in the process of trying to get some kind of statement from them as to why they downclock for CUDA and OpenCL.

I will pass on anything they tell me.

---

My limited experience with Nvidea tech support was excellent.

---