So I thought my points per day were low, currently getting around 21k for the following hardware :

GTX 970
Intel 4600 GPU
3x Intel CPU cores from my 5930k CPU

So I checked my tasks for my main machine containing the 970 and the 5930 and the results look a little odd :

http://setiathome.berkeley.edu/results.php?hostid=7407076

1308 tasks but only 256 are valid?

Any ideas whats going on here? seems like a lot of wasted GPU time :(

---

I also do a little einstien at home on this GPU as well, just checked the results on there page and nearly all the results are coming out as invalid :s

http://einstein.phys.uwm.edu/results.php?hostid=11700141

---

Ok I checked this thread :

http://setiathome.berkeley.edu/forum_thread.php?id=76484

And noticed I was also getting cuda 32 units, I thought this might be because I was running multiple units on one card, so I deleted mu app.xml file which I had edited and now its just downloaded cuda42 units, are these right for my card?

---

Ok ran the lunatics installer and that seems to have sorted it.

Question though, why on a default install did it start crunching cuda_42 units? I bet there are a load of users out there crunching the wrong units on the wrong cards because the don't know any better and just installed it and left it?

Also how often are new units released? i.e how often will I need to run the latest version of the lunatics installer on my machines to make sure I am running the correct units?

Is this just a Nvidia issue? work units seemed about right when I was just using AMD cards.

Oh and lastly how can I now run multiple units on my card? looked at the new app_config.xml and there is a lot in there now.

---

Update, Checked my Einstin stats and all the units say failed, just run some Milky way units and they are all coming up and validation inconclusive

Is there issues with 970's at all??

It seemed to do Seti fine when it was in my other PC?

---

No problem with my 970 that I'm aware.

Ryan,


Most of your invalids are from Cuda 32 work units.

Rerun Lunatics and install the cuda50.

That is the best option for the 970.

Any cuda 32 that you downloaded will still be in your cache until you burn thru them all. Once you are done with those the cuda 50 (or 42 since you installed those) will download and process.

I would recommend you change it to the cuda50.

Once those start to process we will see how they do


Zalster

edit..

make sure you download the drivers from Nvidia directly and not from Microsoft. Always do a clean install

Yes I am only running cuda 50 now, all that have uploaded are fine apart from one thats inconclusive but I hope thats a blip.

All my milky way units are coming back inconcusive now though and the last lot of einstien I did all came back invalid, will post over there for help but I wonder if its all related?

---

Are you overclocking? or changed any of the setting of the GTX 970?


edit...

I looked at that inconclusive you mentioned, it is part of the set work units that are all invalid due to a problem with the server so ignore that one

Nah its running as it was out of the box, it is factory overlocked though (EVGA FTW), never gets hot though.

Just testing a few things (underclocking / speeding the fan up)

However it seemed to work ok in my other box, cant see any errors on that machine.

---

Rerun Lunatics and install the cuda50.

That is the best option for the 970.

Any cuda 32 that you downloaded will still be in your cache until you burn thru them all. Once you are done with those the cuda 50 (or 42 since you installed those) will download and process.

No, that's not how Lunatics works - as regular readers will know.

Once you have installed the Lunatics applications - agreed, cuda50 is right for this card - every task in the cache will be run using the cuda50 application. Some of them may still have a cuda32 or cuda42 label attached, but that has no effect on the processing.

Any thoughts as to why the 32's and some of the 42's kept failing? and why the other projects are failing as well?

---

I'm having the same issue on two machines, this one (Win32) and this one (Win64). They are both generating hundreds of invalids.

* Both of them have GTX980s. (Hope this gets fixed or I've just wasted a lot of money. :^p )
* This happened running 100% stock on a fresh install of Windows with only the latest NVidia driver set (347.09), and continued once Lunatics with MB CUDA50 was installed.
* As indicated above, 32-bit/64-bit OS doesn't seem to make a difference.
* It doesn't seem to affect CUDA AP... the results are validated.
* It didn't affect E@H CUDA when I was running that. This host had zero invalids.

Here's a comparison of the bad result from mine and the good result from someone else of the same MB CUDA work unit. The others I checked also shows these excessive results. From watching them, they complete (sometimes long) before reaching 100% of course.

So this seems to be just an incompatibility in MB CUDA on Maxwells which is part of the stock code and hasn't been altered by Lunatics yet.

Hope this is of some use and if I can help out with any testing, etc. if there's a fix in the works please let me know... one of those machines is just a cruncher so I don't mind breaking it for a good cause. :^)

---

how much CPU are you giving each MB?

how much CPU are you giving each MB?

0.4 on both. GPU util. usually about 78-80%.
(Edit: running with count=0.5 as they aren't very powerful.)

---

Ryan,

post your app.xml file here so we can look at it. Just want to check something.

Zalster

OK, I moved both 980s to my main machine. In doing so I noticed the elephant in the room that seems to have been overlooked:

Power.

I had one on a 525W named PSU and another on a 680W no-name, but I bet neither one could deliver the sustained ripple-free 14A the cards need. (Also NVidia has really pushed the reduced power requirements of the Maxwells... maybe to the point of "exaggeration.") Now that they are on the 1000W Azza Titan with the 86A 12V rail, all the work units are completing.

What make/model of PSU do you have Ryan?

---

I think I have part of the answer here. It's just a matter of the server not taking no for an answer. I am the 8th one to get this work unit and in about 2 days it will be crunched and I will get no credit for it.

---

I think I have part of the answer here. It's just a matter of the server not taking no for an answer. I am the 8th one to get this work unit and in about 2 days it will be crunched and I will get no credit for it.

I think this is a different issue... that one is a bunch from one day. This issue is Maxwell 970/980 cards generating 80% invalids from any time.

I guess the MB CUDA just stresses the GPU more than AP or the Einstein code.

---

I too am getting lots of invalid results is this because there is a problem with my fan when I switch on my machine. Have to log in then out again as the fan is making a terrible noise then restart my machine. Think I need a new one even though I have not had it a year it has been into the shop because the P.S.U. went down and now the fan

---