It's disabling the driver watchdog completely. It's supposed to stop errors from happening.. but if there's a real error you'll have to go for the hard reboot.

Or was it increasing the watchdog timeout? Here SK suggested to put 20s tolerance in there default 2 s). I've tried it with 5 s and got a froozen screen for ~20 s (upon suspending a NOELIA) followed by the usual driver restart. Instead of trying also with the 20 s setting I reverted the change completely.

MrS

---

Scanning for our furry friends since Jan 2002

Carlesa25 wrote:

It seems that I have a problem on Linux / Ubuntu with the GTX 770 and Noelia tasks, performance is pitiful salary at the GPU and no CPU usage.
I can not get off to a less than 319.23 as driver do not support the 770.
Is there any forecast of when this issue will be resolved or have to wait for a new driver from Nvidia?

Well, it's obviously a driver issue, since it works with the older versions. I can't see anything BOINC or GPU-Grid could do about this other than to inform nVidia and hope they'll fix it at some point. If the most recent beta drivers are still not working, chances are that nVidia doesn't yet know about this problem.

As a work around you switch the GTX770 to a windows box, if you've got any.

I suggest a different workaround:
As the GTX 770 is basically a GTX 680 with higher clocks, the GTX 770 should work with the previous drivers, if you include the appropriate line into the nv4_dispi.inf file (this method is for Windows only, so a Linux guru should tell us how to do it under Linux)
You should look for:
%NVIDIA_DEV.1180% = Section021, PCI\VEN_10DE&DEV_1180
You should copy the whole line below the original (the section number may be different), and then change both 1180 to 1184:
%NVIDIA_DEV.1184% = Section021, PCI\VEN_10DE&DEV_1184
and then you should look for:
NVIDIA_DEV.1180 = "NVIDIA GeForce GTX 680"
copy the whole line below the original and then change 1180 to 1184, and 680 to 770:
NVIDIA_DEV.1184 = "NVIDIA GeForce GTX 770"
After these modifications the old driver should recognize the new card.

so far the only reports of this happening have been from SK and TJ. Are you guys just watching more closely than others.. or is the error only happening on your systems?

Yesterday, I caught one that had been running for 10.5 hours and was at 0.00%, GPU at 99%, memory controller at 0%, I aborted it and it was a first run (ended in 0).

Hello: I do not understand is that with short tasks work perfectly my GTX770 (Ubuntu 13.04) and the long NOELIA failure, so far I have not tried with a different long task.

I think the problem is in Noelia tasks.

Moreover Einstein I finished seven tasks perfectly, no problems. Greetings.

Einstein tasks are entirely different, and al an older version of CUDA.

Einstein tasks are entirely different, and al an older version of CUDA.

That's true, but Einstein, Albert, Milkyway are nice projects to test your setup and drivers. Most WU' run fast so you can see what happens.

I agree with Carlesa25 and others that these set of Noelia's have strange behavior. I use driver 320.49 and 320.18 for Noelia's and most finish, but in longer time or error quite immediately. The error rate is higher. I don't think a driver change will resolve all problems, and Linux is more efficient than Windows when crunching.

---

Greetings from TJ

Such a pity about the NOELIA tasks: wasting energy and computer resources.....

John

Hi Noelia tasks in Windows 8-64bit are working well, so far, on the GTX770, the GPU load 86% + 22% CPU.

In Ubuntu keep failing.

Three people reported the 1% MCU load, and I changed drivers just in case they were the issue. You have to be using GPUZ to see the MCU load, it’s not listed in Precision/Afterburner. I suggest anyone seeing a task run for ages check this in GPUZ.

I didn’t disable the driver watchdog, just give WU’s a bit more time before the system restarted the drivers. Prior to doing this, I had driver restarts for most task types and higher error rates. I think it has fixed some issue but not others. I still get driver restarts, but they are rarer. Occasionally I get some screen lag, but I’m now using the iGPU for display.

While the 319 drivers might have issues, I don’t think the drivers are to blame with these WU’s. The GTX770 would probably work with older Linux drivers. I’ve done this before on Linux and Windows and Boinc recognised the cards correctly in both cases.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Hello: I'm using EVGA-Precision and gives the same readings on all parameters that GPU-Z. 320.49 driver.

At present after two hours of work have you done on 27% of the task, which is fine.

I don't think EVGA precision gives a Memory Control load reading, so you have to use GPUZ to get that.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I don't think EVGA precision give a Memory Control load reading, so you have to use GPUZ to get that.

SK is right, you can't read the memory controller load in PrecisionX, you need GPU-Z or GPU-Shark to see the controller.

I don't think EVGA precision give a Memory Control load reading, so you have to use GPUZ to get that.

SK is right, you can't read the memory controller load in PrecisionX, you need GPU-Z or GPU-Shark to see the controller.

Hello: I have installed and use, GPU-Z 0.72 and EVGA-PrecisionX 4.2.0.2143, and the same sensor readings and see one another, do not really understand what you say.

GPUZ
Sensors Tab,
Memory Controller Load.

~28% is normal for a GTX660 and 37% for a GTX660Ti, but some WU types are a bit higher and some a bit lower. If the load is 1% its really bad!

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

That's correct Precision X does not show MCU load or GPU power, but EVGA NV-Z does.
I use that most of the time.

---

Greetings from TJ

Hi Noelia in Windows 8 Task completed smoothly.

It took 28,675 sec. almost like the GTX680 that I could compare.

Deputy screenshot with GPUZ readings and EVGA, which give the same data.

https://picasaweb.google.com/lh/photo/-bwUatfEBByRcJpZTfJikdMTjNZETYmyPJy0liipFm0?feat=directlink

I was going to keep my GTX 650 Ti with 1GB memory on Longs even with the occasional crash, but it has occurred to me that this might give the wrong feedback to GPUGrid. That is, if they get 5 errors (or whatever their limit is), they might conclude that it is a bad work unit and discard it, when in fact it is merely due to 5 cards with not enough memory. Therefore, I am going to Shorts, but also enabling Beta testing, so they can try out their stuff (if they choose to) before releasing it. (I make sure my card is running at the default Nvidia chip speed rather than the card factory overclock, since a test of how unstable the card is does not do them any good in evaluating their work units).

But if they knew about the larger memory requirement, I think they should have communicated that to us in advance to save us all a lot of trouble. And if they didn't, then that is what the Betas are for.

Ah, Carlesa. You leave K-Boost off I see. That's why it shows your FB %. Than yes, they do show the same data.

I've tried the latest (320.49) driver on my least reliable host (it has a WinXPx64), and it became completely unreliable :) Every (Noelia) task is stuck at 0% with 0% GPU usage using the 320.49 driver (btw it's CUDA 5.5).
I'm reverting to the driver downloaded by the Windows Update (307.9)

I've tried the latest (320.49) driver on my least reliable host (it has a WinXPx64), and it became completely unreliable :) Every (Noelia) task is stuck at 0% with 0% GPU usage using the 320.49 driver (btw it's CUDA 5.5).
I'm reverting to the driver downloaded by the Windows Update (307.9)

Hello: In Windows 8-64bits I am using the 320.49 without problems, both short tasks like Noelia.