Seti@Home with Titan ▼

[Uzume]

Just got a Titan ▼ in for my Seti box (Chibineko in sig) to replace one of the Vega 64 LC's which has gone to GPU Valhalla. Currently running the Titan ▼ at stock and the remaining Vega 64 LC at a memory only OC of 1020MHz. Using standard Boinc client but not in screensaver mode of course. Boinc uses 6 of the 1700X's 8 cores, so no GPU starving.

 

The Vega 64 LC's have been averaging 5-7 minutes per work unit and the new Titan ▼ is sitting at around 20 minutes per work unit. Rather underwhelming. A quick screenshot example is below. Currently this card is severely under-utilised by Seti@Home.

 

Just a heads up.

 

[alextheaverage]

I know in Primegrid you can add some parameters to the config to run multiple units on one GPU, I saw people doing four on their R9 290's due to utilization problems.  The tasks sometime end up running slower individually but over all the GPU gets pegged at a full 100% so people are able to complete many more tasks in a given time frame.  It's not like Seti@home has super huge four to eight day GPU tasks like Primegrid but if you can run multiple tasks you could put out a ton on that beast.

[Uzume]

On 07/02/2018 at 1:39 PM, alextheaverage said:

I know in Primegrid you can add some parameters to the config to run multiple units on one GPU, I saw people doing four on their R9 290's due to utilization problems.  The tasks sometime end up running slower individually but over all the GPU gets pegged at a full 100% so people are able to complete many more tasks in a given time frame.  It's not like Seti@home has super huge four to eight day GPU tasks like Primegrid but if you can run multiple tasks you could put out a ton on that beast.

 

Thanks for that. It took a day of messing around with the app_config file but I found I could throw 3 work units at each card (Titan V and Vega 64 LC) simultaneously without major penalty per work unit. I don't even think this is pushing the Titan V but I will wait till I can isolate this card in a single GPU system before trying to push it to its limits.

 

This will greatly increase throughput.

[alextheaverage]

That's awesome!  It's a shame that the cards aren't properly utilized in the first place, but then again on a Titan V a Seti@home task would take like 30 seconds to a minute.  Tripling throughput sounds really nice though.

[FloRolf]

Do you mind testing some F@H as well @zzrhardy? Would be quite interesting! 

[alextheaverage]

15 minutes ago, FloRolf said:

Do you mind testing some F@H as well @zzrhardy? Would be quite interesting! 

overclock.net has a full chart of PPD figures that are user submitted, multiple data points make for a great way to get a real idea of how multiple skews of GPUs run.  The site seems to be down for me for some reason so I can't link it at the moment.

[FloRolf]

21 minutes ago, alextheaverage said:

overclock.net has a full chart of PPD figures that are user submitted, multiple data points make for a great way to get a real idea of how multiple skews of GPUs run.  The site seems to be down for me for some reason so I can't link it at the moment.

Yes, I know. I always use this right here for comparison, but there's no Titan V in there yet. 

Thats why I am asking

 

https://docs.google.com/spreadsheets/d/1vcVoSVtamcoGj5sFfvKF_XlvuviWWveJIg_iZ8U2bf0/pub?output=html#

[Uzume]

I have installed Folding@Home on that box - just letting it run for a day or so to get some accuracy.

 

Edit: Results below.

 

Man - I thought S@H was clunky, but F@H made it look advanced. It failed to use the Vega 64 LC at all and it doesn't seem to be able to properly utilise the Titan V either.

 

 

Edited February 9, 2018 by zzrhardy

[alextheaverage]

16 hours ago, zzrhardy said:

I have installed Folding@Home on that box - just letting it run for a day or so to get some accuracy.

 

Edit: Results below.

 

Man - I thought S@H was clunky, but F@H made it look advanced. It failed to use the Vega 64 LC at all and it doesn't seem to be able to properly utilise the Titan V either.

 

 

How long did you let it run for?  Sometimes Folding@home takes up to 3-4 hours to get a decent estimate.  Like how Boincs estimates aren't usually close either.

[Uzume]

I ran it for about 8 hours. Was going to leave it running over night except it wasn't even attempting to use the Vega 64 LC so it would have been a waste. 

 

Edit: maybe it didn't like being in a mixed Radeon/Nvidia compute box.

Edited February 10, 2018 by zzrhardy
clarification

[FloRolf]

On 9.2.2018 at 12:02 PM, zzrhardy said:

I have installed Folding@Home on that box - just letting it run for a day or so to get some accuracy.

 

Edit: Results below.

 

Man - I thought S@H was clunky, but F@H made it look advanced. It failed to use the Vega 64 LC at all and it doesn't seem to be able to properly utilise the Titan V either.

 

 

47000 points per day is literally nothing. My 1080 makes roughly 800000. Weird though, as 8 hrs should be enough for it to settle. 

[Ithanul]

21 hours ago, FloRolf said:

47000 points per day is literally nothing. My 1080 makes roughly 800000. Weird though, as 8 hrs should be enough for it to settle. 

Could be no passkey was added.  Without a passkey, the points will look lower on any card.

 

Still, most likely most BOINC projects and F@H are not coded to take advantage of a Titan V considering hardly any of the member base would have such a card.  Also, F@H been on the core 21s for quite some time which came around the Maxwell era (since Pascal is a refined Maxwell, it still runs great on those).  At most, a new core would have to be coded up to take advantage of new hardware on the Titan V.

 

Re looks back again, I could see one issue.  CPU is at 14 out of 16 logical cores from the 1950X.  At most, that needs to be knocked down to 12 to make sure both GPUs are feed and the OS has a core along with the fact to keep it off an odd number (F@H has this funky behavior with odd count).

[Zalster]

Your times for that Titan V on Seti look way off.  You aren't showing what applications you are running for that work unit. If I had to guess based on time to complete, I'd say you are running a cuda 3.2  You should repeat the test and at a minimum run the cuda 5.0 application or better yet run the OpenCl SoG.   Ithanul is correct that you should be devoting 1 core to each GPU, especially for the OpenCl.

 

There was another Titan V on Seti about a week after it was released. His times to complete were 1 minute 23 seconds I believe. Comparable to the Special Cuda 9.2 application from Petri. 

[kurtstir]

On 2/9/2018 at 5:02 AM, zzrhardy said:

I have installed Folding@Home on that box - just letting it run for a day or so to get some accuracy.

 

Edit: Results below.

 

Man - I thought S@H was clunky, but F@H made it look advanced. It failed to use the Vega 64 LC at all and it doesn't seem to be able to properly utilise the Titan V either.

 

 

Hey anychance I can get an update on your folding speeds on the correct tab? You accidentally showed us your cpu tab

[Uzume]

Sorry bout that, I am unfamiliar with the folding software. That box is no longer a mixed nVidia/ATI setup so I just resinstalled the FAH client.

 

It appears to be using both V's so that is a good step, but will need for it to settle down for at least half a day to get any halfway meaningful numbers.

 

Edit: also the other box with 3x V64s is also folding on all cards. The original problem definately seems to be mixing team Green and Red in a single box.

[Uzume]

Been going a few hours now - here's a preliminary peek, first the 2x Titan V box with 10 of 16 threads CPU used:

 

 

And here are the underwhelming 3x Vega 64 box with 24 of 32 threads used:

 

 

It was hovering around 1.4M PPD for a while then settled on the 800K figure as it changed WU types over time.

 

I'll only update this post if the figures change markedly after 12 hrs or so.

 

[GOTSpectrum]

11 hours ago, zzrhardy said:

Been going a few hours now - here's a preliminary peek, first the 2x Titan V box with 10 of 16 threads CPU used:

 

 

And here are the underwhelming 3x Vega 64 box with 24 of 32 threads used:

 

 

It was hovering around 1.4M PPD for a while then settled on the 800K figure as it changed WU types over time.

 

I'll only update this post if the figures change markedly after 12 hrs or so.

 

My Vega 64 produces 600k per day. So yours does seem lacking.

[BuckGup]

Do you enter your passkey? I made 120K with a 6800K and 1080 then entered passkey and make over 1 mil now

[Uzume]

6 hours ago, BuckGup said:

Do you enter your passkey? I made 120K with a 6800K and 1080 then entered passkey and make over 1 mil now

 

Yes - it had a passkey inputted. Probably one of a dozen other possible issues tho because the F@H user interface as created by Satan himself.

[GOTSpectrum]

On 22/05/2018 at 4:05 AM, zzrhardy said:

 

Yes - it had a passkey inputted. Probably one of a dozen other possible issues tho because the F@H user interface as created by Satan himself.

Your log generally has lots of information in it with regards to troubleshooting.  

[GOTSpectrum]

btw @zzrhardy this is my VEGA 64 currently.