My new GTX 980s are supposed to run memory at 7000MHz, but they are running at 6000MHz. When using EVGA Precision 4.2.1 to increase the memory speed by 500MHz, when I click Apply, it goes right back to 6000, and the memory speed slider goes right back to +0. If I do this on my other machine, with GTX 780s, the memory speed DOES get increased, and the slider remains at +x, where x is the increase specified.

Is there some weirdness with the 980s? I vaguely recall something in another thread, but I can't find it.

Any help would be appreciated; I'd like to get my full 7000MHz!

---

I had much the same with Zota Firestorm - the version that worked for all my other GPUs from various manufacturers refused to control the fan on my GTX980. Upgrading to the latest version (not the box shipped version) was a partial cure - While it will set the fan speed correctly to 100% I have to go in manually and set after a power-off reboot, but its OK after a restart reboot....

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

OK, maybe the latest EVGA Precision will set the VRAM speed. (I'll test it when I get a chance).

But in the meanwhile, why doesn't the card run at 7GHz? it's certainly NOT idling. Normally cards will downclock GPU and VRAM when the card is way underutilized. But this one is running at 50%+ of GPU utilization, so you would think it wouldn't, yes? And GPU speed is way up at boost levels (or a little beyond), so why isn't VRAM?

---

Is this the thread you were thinking about?
http://setiathome.berkeley.edu/forum_thread.php?id=76997&postid=1657723

I see much the same on my ATI card.

Most of the time it spends at the minimum speed, then it gets a task(s) that wants more and it kicks up to requested speed.

I'm not exactly sure what request pushes it up, but I can definitely see it happen. I'm guessing it is when it finds something interesting and 'looking deeper' into it, but who knows.

---

Here's the Truth, direct from EVGA support:

"As soon as a CUDA or OpenCL task is started, the card is limited to "P2" as the highest performance state. This is meant for some less-demanding load and limits GPU memory clocks to 1.5 / 3.0 / 6.0 GHz on GM204 cards, depending on how you want to count DDR clock speeds.

Regular tools like EVGA Precision, Afterburner etc. only set values for the P0 state. If you increase the memory clock there, it applies just fine to regular 3D loads like Heaven, but has no effect on the memory clock in P2. Strangely a GPU OC set in those tools applies to both, P0 and P2, so this works just as expected.

But you can use nVidia inspector to set the memory clock speed for P2 separately."

Quoted from this thread
http://einstein.phys.uwm.edu/forum_thread.php?id=11044

It would seem that the gtx 900 series cards (GM204) operate differently with cuda computation tasks. The card is working as intended with no flaws in the design of the memory controller. Please let us know if there is anything else we can assist you with in regard to your EVGA products by contact support at support@evga.com or 888-881-3842 for further assistance.

Thank you,
EVGA

Note that the part I bolded seems to contradict the rest of the statement. I will try to follow up and find out what I can. But at least it's a start.

---

Is this the thread you were thinking about?
http://setiathome.berkeley.edu/forum_thread.php?id=76997&postid=1657723

Yes, it is, thank you very much.


Nvidia Inspector it is for that machine.

---

Jravin,

I know the process has been detailed in that thread. Just in case though, make sure you first exit BOINC, then if you want to set your P2 clock to anything higher than the stock rated 3500mhz make sure you do that in the AFterburner or Precision utility, such as a +50, +100 mem clock offset. I'm sure you get the picture. Then will you be able to go into the Nvidia inspector and set the clock to that maximum defined speed you set in the OC utility.

Also, make sure you go into the Nvidia control panel and in the global settings set the cards to "prefer maximum performance" rather than the default "adaptive" setting. This singular change will make it so your machines don't reboot randomly.

Hope that helps!

---

Jravin,

I know the process has been detailed in that thread. Just in case though, make sure you first exit BOINC, then if you want to set your P2 clock to anything higher than the stock rated 3500mhz make sure you do that in the AFterburner or Precision utility, such as a +50, +100 mem clock offset. I'm sure you get the picture. Then will you be able to go into the Nvidia inspector and set the clock to that maximum defined speed you set in the OC utility.

Not true, you can adjust your memory clock speed higher than stock directly in NvidiaInspector. You don't have to use the normal game overclocking solutions. First exit BOINC fully, then adjust the memory clock to whatever offset you want for P0 in NvidiaInspector. Apply. Then select the P2 state in the drop down dialog and then slide the memory slider fully to the right to the offset adjusted memory speed. Apply. Do the same for each video card from the card selection dialog on the left. I bump my memory speed +100 for both GTX970 cards directly in NvidiaInspector. Running at 7210 Mhz. Exit NvidiaInspector and restart BOINC. You should see your memory speed running at your higher offset speed. Be careful in the first offset selection process because NI sometimes glitches and the speed offset falls back. Just reset the offset again in P0 and apply. This never happens for the second card for some reason, just for the first.

Cheers, Keith

---

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

A proud member of the OFA (Old Farts Association)

Well so far today I've had the video driver crash at least 3 times with moving memory speed up from 3.0 GHz to 3.5 GHz

Always on the same video card thou. and only when crunching APs only

I'm thinking I might be stressing the system too much with 4 cards trying to maintain that 3.5 GHz.

If it happens again I will have to back off to 3.0 GHz which I'm thinking I might have to do.

I didn't see this with MBs so I wonder if it's only a AP issue?


Edit.. It also seems to only happen when going from completing 1 AP to start another. There is a huge uptick in Kernal usage initally at that time and I think that might be triggering the driver crash.

Zalster

How many AP per card are you trying? Does the video card get enough CPU time to feed it properly? Are you using any tools to monitor the systems resources while an AP task finishes and another one comes aboard? ProcessLasso might be handy. Also, I was just reminded of another handy utility to diagnose system slowdowns. DPC Latency Checker:

http://www.thesycon.de/eng/latency_check.shtml

Might be useful. Of course, if downclocking the memory for the recalcitrant card fixes the problem then problem solved in a timely manner. Could be just a weak card compared to the others.

Cheers, Keith

---

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

A proud member of the OFA (Old Farts Association)

I ended up moving them back down to 3.0 GHz and it all resolved.

As it was, there was stuttering of the screen and lagging so now that I've gone back it's all good again.

I may play with it again and try that program that Jason suggested like you did.

But for now, the old motto goes "If it ain't broke, don't fix it!!"

Well it happen again with the GPUs set at their defaults.

So it's not the OC of the GPUs causing it.

SIVX64 show a huge uptick in Kernal activity at the start of a new AP starting to be crunched so it looks like that is the item that is causing the Graphic Driver to crash.

Ideas people?

The Driver recovers and continues on to crunch ok, so .....

What is the CPU allocation for AP GPU tasks? It sounds like you don't have enough CPU to feed the card. I run a mix of MB and AP tasks on each card but have only seen one case where a card was solely occupied by AP tasks and the system got a little laggy. Increasing CPU to a full core solved that. I think that was when we had the rare case of a dearth of MB and abundance of AP. I've also seen that running ProcessLasso helps smooth out my systems when they are crunching at maximum output and power draw. If you aren't running that yet, you might give it a try. How much red color are you seeing in the core utilization graphs in SIV64? I never get more than one quarter to a third on a core and only for a short while till ProcessLasso rebalances the load to another core to give the original core a rest. I think that keeps the core temps down. I don't have reboot issues. Standard troubleshooting methods apply, make one change to the system at a time and observe. Good Luck.

Cheers, Keith

---

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

A proud member of the OFA (Old Farts Association)

Hey Keith,

I need to try that ProcessLasso.

CPU was 0.35 C for each work unit. It only happens at the onset of a new AP work unit. So my suspicion is that the kernal goes to a near a full core at the start then back down to what it after a few seconds.

But that is enough to crash the driver.

I just downloaded that program and will see how it response.

This only ever seem to happen with the APs, not the MBs

Thanks

Zalster

ProcessLasso is a pretty nifty utility. It's got an overwhelming amount of configuration options but I pretty much leave it configured at stock. Just make sure you use the ProBalance option. The object of the program is to maximize system responsiveness at all times no matter what programs or intensive processes are running. I agree, that the initial startup of an AP task is what really sucks down a system's responsiveness. I think what happens is that the heartbeat for a running task and also the heartbeat for the video driver doesn't get quick enough attention and the system decides it best to down and restart the video driver. I've seen enough cryptic messaging to at least suggest that is the case. I don't have a good enough understanding of Windows design to really comprehend what is going on when BOINC has an upset or the video driver resets. Just that ProcessLasso has most definitely helped the system responsiveness and stability. My $0.02.

Cheers, Keith

[Edit] Forgot to state that I used to run with a full CPU core for each AP task but then was able to reduce that down to 0.5 CPU core for an AP task with no consequences other than I was able to run more concurrent CPU tasks and still maintain a stable system.

---

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

A proud member of the OFA (Old Farts Association)

Hi Folks,
Someone posted about a new Nvidia hotfix driver on release. But whomever didn't have access to a copy.

Well its available for d/l via Techpowerup website:-)

Just in case some brave soul wants to have a go, apparently it includeds Cl 2.1
or something:-)

Regards,

---

Cliff,
Been there, Done that, Still no damm T shirt!

Hi Folks,
Someone posted about a new Nvidia hotfix driver on release. But whomever didn't have access to a copy.

Well its available for d/l via Techpowerup website:-)

Just in case some brave soul wants to have a go, apparently it includeds Cl 2.1
or something:-)

Regards,

Jamie has already tested it at Beta, and the current NV APv7 app there doesn't work with it, fresh app going is to be released with a fix.

So don't upgrade to the Hotfix if you're running AP on a NV GPU, and you don't have the skills to change easily to a fixed app.

Claggy

Hi Claggy,

Don't worry I had NO intention of installing the hotfix driver until I knew it was safe to use.

NVidia has released to many DC busting drivers in the past to simply jump in and install any of their drivers.

Regards,

---

Cliff,
Been there, Done that, Still no damm T shirt!

You know, I have been thinking this whole thing over (as OP) and I have come to the conclusion that Nvidia is screwing the people that bought these cards by not allowing them to run full bore in all situations.

It's like if you bought a car and the manufacturer didn't tell you it had a governor so that you couldn't take it out on a race track and get it to run all out.

Pretty cheesy I think. Why in Hell would they do this? Are they trying to protect the market for Titans? Do the Titans suffer from the same throttling? America wants to know!

Right now, my 980s are significantly underperforming the 780s on my other cruncher. I wonder if Nvidia is going to correct this soon; they could very well have a lawsuit on their hands for this. It is MUCH worse IMO than the earlier flap about using the last 1/2 Gig of RAM on the card.

---