Latest Nvidia drivers 331.40 & Titans
 

Anyone trying out the latest 331.40 driver on a Titan?

I'm seeing some major slowdowns.

I'm predominately doing P-1, and I've gone down from 80GHz-days/day (ish) to about 50 (ish).

-- Craig

I'll neither confirm nor deny the double-precision config item checked in the driver was cleared. :devil:

It must have cleared on installing latest driver.

Tip for a new Titan player - when doing LL/P-1 on your Titan, make sure the Double-Precision config item is checked. :)

-- Craig

So performance turned out about the same, then? Is that a WHCL or a beta driver?

I wish I had more accurate stats. But there's no noticeable performance difference.

-- Craig

[QUOTE=nucleon;355311]I wish I had more accurate stats. But there's no noticeable performance difference.

-- Craig[/QUOTE]

A general impression is usually as far as I go, anyway. Thanks! :smile:

I installed latest beta driver 331.40, as I had 320.59 which I believe came with Cuda 5.5. In the 320.59 driver I had to drop memory hz to 2500 (-500), as one of the first issues with titan was memory heating up on the back side of the card where there is no fan. I was waiting for that to get fixed, and it seems like new drivers has fixed some of this. Did Nvidia only activate memory inside the card unless more memory is needed ? Mabye something changed with gtx 780 release.

To get these results, I am now at 3004 mhz on mem clock. 940mhz on gpu clock. EVGA precision power set to 106%. This has now run stable overnight.

- Driver version: 331.40
- Mem clock: 3004 mhz
- GPU clock: 914-940 mhz
- Compute 35 cuda 64bit cudalucas on windows 64bit
- EVGA power set to 106%
- test exponent: 38000009

Titan with DP switch on:
[CODE]
Iteration 10000 M( 38000009 )C, 0xfd9116e3760e4571, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:16 real, 1.5878 ms/iter, ETA 16:45:21)
Iteration 20000 M( 38000009 )C, 0xd91fe21f272e5099, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:15 real, 1.5032 ms/iter, ETA 15:51:29)
Iteration 30000 M( 38000009 )C, 0x7ec1302ed5173c26, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:15 real, 1.5025 ms/iter, ETA 15:50:51)
[/CODE]

Titan with DP switch off:
[CODE]
Iteration 10000 M( 38000009 )C, 0xfd9116e3760e4571, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:22 real, 2.2743 ms/iter, ETA 24:00:02)
Iteration 20000 M( 38000009 )C, 0xd91fe21f272e5099, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:22 real, 2.1882 ms/iter, ETA 23:05:08)
Iteration 30000 M( 38000009 )C, 0x7ec1302ed5173c26, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:22 real, 2.1871 ms/iter, ETA 23:04:04)
[/CODE]

I ran with these settings overnight, and they seem to run stable. (I did not have these settings run stable on driver 320.59)

[QUOTE=Manpowre;356727]Titan with DP switch on:
[CODE]
Iteration 10000 M( 38000009 )C, 0xfd9116e3760e4571, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:16 real, 1.5878 ms/iter, ETA 16:45:21) ....[/CODE]Titan with DP switch off:
[CODE]
Iteration 10000 M( 38000009 )C, 0xfd9116e3760e4571, n = 2097152, CUDALucas v2.03 err = 0.2090 (0:22 real, 2.2743 ms/iter, ETA 24:00:02) .... [/CODE][/QUOTE]Thx for lists, i´m looking a long time for this comparison. But only 33% shorter ETA? DP is nearly 10times faster than the other GFORCES.

[QUOTE=MikeBerlin;356737]Thx for lists, i´m looking a long time for this comparison. But only 33% shorter ETA? DP is nearly 10times faster than the other GFORCES.[/QUOTE]

Dunno. I think I saw similar timings earlier with dp switch off.
I dont think it is 10 times though. We are talking 7.5b transistors even with dp switch off.

[QUOTE=MikeBerlin;356737]Thx for lists, i´m looking a long time for this comparison. But only 33% shorter ETA? DP is nearly 10times faster than the other GFORCES.[/QUOTE]

well if you take original post: 50*100/80 = percentage of with DP = 80ghz days.. then its 62ish.. since original poster reinstalled driver, and DP setting was off, he got 50ish instead of 80ish,, which seems like ishish same result as I get something like 33 percent difference..

I guess there is complexity in FFT that makes this 33% instead of 1/3 stepper with DP compared to 1/24 for 680/690/780/790 series of geforce.

then again, the difference is on Titan only with and without DP setting. For 6xx platform and 7xx platform there is really no dp on/off setting, and these platforms also got less transistor counts inside GPU, so there is huge difference with cudalucas compared to Titan with its 7.5 bill transistors and 2600 ish computecores.

A WHQL 331.58 Forceware is out, anything regarding it?

[QUOTE=Karl M Johnson;356968]A WHQL 331.58 Forceware is out, anything regarding it?[/QUOTE]

I tested 331.58. I did 2 tests on a Titan:
* FFT length tests 3.5m-4.5m 1024 step, DP on. marginal differences. Some fft lengths marginal quicker, some fft lengths marginal worser...
* Exponent test. exactly same time, and timings as with the 331.40 beta driver.

so I dont see the big differences.. I just hope it is stable...