Have the scientists/developers/programmers been able to make any progress on updating the apps for GTX Titan compatibility?

I've got mine busy generating rainbow tables at present and would like to put it back to work here at GPUGrid.

Operator

---

Thought I'd add something I noticed today. The details page for my system with the GTX Titan says it has 4GB of ram.

"Coprocessors NVIDIA GeForce GTX TITAN (4095MB) driver: 314.9"

The problem is that the GTX Titan actually has 6GB of ram.

Could that account for why tasks are failing? A memory size mismatch between what the system thinks the card has and what it actually does have?

Has anybody else with a GTX Titan noticed this discrepancy?

Operator[/b][/quote]

---

This is from the BOINC log at startup:

3/11/2013 5:56:31 PM | | NVIDIA GPU 0: GeForce GTX TITAN (driver version 314.9, CUDA version 5.0, compute capability 3.5, 4096MB, 8384512MB available, 4989 GFLOPS peak)

3/11/2013 5:56:31 PM | | OpenCL: NVIDIA GPU 0: GeForce GTX TITAN (driver version 314.09, device version OpenCL 1.1 CUDA, 6144MB, 8384512MB available)

Strange. Only the OpenCL section is reporting the correct memory amount.

Operator

---

The NVIDIA driver API seems only to return a 32-bit value.

It's been reported.

I saw over at the SETI forums that the Titan memory error was reported to the BOINC alpha list.

So I guess the BOINC devs are working that.

But are you indicating that would be a Nvidia driver problem instead?

Operator

---

Well. In the first place, those who run a 32-bit operating system will only ever see 4GB - that's the limit.

For 64-bit OSs: you'll have seen Claggy's reports of the 'wacky memory' bugs, when BOINC was reporting millions of millions of megabytes for some cards. The BOINC devs can't afford to buy a new card to test every time they come out - certainly not at Kepler and Titan prices - so Rom remoted in to my GTX 670 to work out what was going on.

Apparently, the 64-bit drivers require the programmer to pass a 64-bit memory buffer to return the memory size: but they only populate the lower 32 bits. Any initialisation garbage in the upper 32 bits remains unchanged - that's what led to the wacky memory sizes. I think Rom cured that by zeroing all 64 bits before use: if the drivers are still only filling in 32 of them, there's not much more to be done.

I wonder what the Tesla drivers do? Anyone got one they can lend to BOINC?

I did fail to mention that I was running Win7x64. Sorry.

I have seen several event log postings over on the SETI board showing the same 4096mb memory amount (you were on that conversation as well).

I have a support ticket in with Nvidia and am in the process of a dialog with them now.

Operator

---

Coming soon, once the Noelia WUs are working ok.

MJH

Hi,

I have a titan and also I do work for other projects.

We found having the default memory speed caused too much errors.

When we clock down the memory from 3000MHz down to 2500MHz, the apps were a _lot_ more stable.

At 3000MHz, some apps would die within mins (4hrs was the best stability). Drop the speed down, stable as a rock. Ran for 48hrs+ no dramas, all results double checked aok.

Hope this helps.

-- Craig

Hi,

I have a titan and also I do work for other projects.

We found having the default memory speed caused too much errors.

When we clock down the memory from 3000MHz down to 2500MHz, the apps were a _lot_ more stable.

At 3000MHz, some apps would die within mins (4hrs was the best stability). Drop the speed down, stable as a rock. Ran for 48hrs+ no dramas, all results double checked aok.

Hope this helps.

-- Craig

Which other distributed computing projects are Titans currently working? do they well on them?

Thanks

Mem clocky are ok at 3000Mhz no worries.

You can check milkyway (with DP enabled) and run 5 tasks in parallel. Primegrid sieve can be ok also.

Mem clocky are ok at 3000Mhz no worries.

You can check milkyway (with DP enabled) and run 5 tasks in parallel. Primegrid sieve can be ok also.

Thanks, I've just read your posts in Milky, 300kppd, it does not seem that impressive, does it? equal or lower than a hd 7970 if i've seen it correctly.

@nucleon: that sounds almost like you could (should?) RMA the card. MAybe try to expose repeatable errors with some of those "artifact tester" utilities at stock clocks. Errors should also pop up there if applications crash.

@Milkyway: yeah, 300k RAC is massive for a nVidia, but not all that great. I remember scoring 260k RAC with my HD6950 maxed out (unlocked shaders, OC, 3 concurrent WUs).. back in the days before POEM came along.

MrS

---

Scanning for our furry friends since Jan 2002

I now have two Titans (as does nucleon I think)and am having no difficulty running Einstein@home (GPU only tasks), as well as the DSTRGEN project.

Haven't tried PrimeGrid but good to know it works.

I have a ticket in with Nvidia on the memory reporting issue (4096mb reported vs 6144mb actual). Nothing back from that yet. The rep said that he was unaware of any issues in BOINC or GPUGrid so I pointed them toward the forums where the problems were mentioned and sent in some screenshots and they are presently "researching" the problem.

So since MW works as well, I may go back over there for a while then.. until we get the longs sorted here.

I don't think I'll be downclocking the memory just yet, unless somebody can conclusively state that doing so fixes the long work units here at GPUGrid.

Operator

---

Milky is open_cl so the same power as top ati is a little miracle. I have not expected that.

DistrRtGen can crunch one task in 530s with titan, overclocked titan in 480s. This is a little better then top ati.

But I'm looking forward to gpugrid. I have buy this card for gpugrid.

But I'm looking forward to gpugrid. I have buy this card for gpugrid.

Same here my friend....

Operator

---

There is mention of a specific OpenCL fix for TITAN in the 314.21 beta driver release notes:

[GeForce GTX TITAN]: Application crashes or a TDR occurs when running
applications that use OpenCL. [1237211]

We don't use OpenCL here, of course - we use CUDA - but it might be worth a test.

I don't know what progress is being made on GTX Titan compatibility with GPUGrid apps, but as has been stated here and in other areas the Titan seems to be happily crunching away on Einstein, Milky Way, Prime Grid, Distributed Rainbow Table generator, etc.

Comments made by Mr. Haselgrove regarding the cause of the misreporting of the total available memory (4096mb vs 6144mb) due to the driver API leaves me wondering if this is in Nvidia's lane, BOINC's lane, or both, and whether anybody is working to resolve this issue or if this issue even matters when crunching CUDA apps with a Titan.

Does this memory reporting mismatch affect how the apps run (or fail to run) on a Titan? Or is it a 'red herring'? I don't know. But obviously some projects are not having any issues at all running on the new card.

Does anyone have a timeline, idea, clue, or rumor regarding when the GTX Titan might be able to successfully support GPUGrid project apps?

Thanks,

Operator

---

Does anyone have a timeline, idea, clue, or rumor regarding when the GTX Titan might be able to successfully support GPUGrid project apps?

Now we have Titans in the lab I'll be doing it late next week.

MJH

Great news!

Thank you Sir!

Operator

---