NVidia GTX 780 Ti

A

AndyE

Limp Gawd
Joined
May 13, 2013
Messages
276
Got my GTX 780 Ti today.
original.jpg


Tried this afternoon to run some folding WU, but it wasn't included in the GPUs.txt file. Thanks to the ff team for quickly expanding the file to include the card. It is now working on its first units.

You will be pleasantly surprised :)

Based on approx 20% of the first WU (P7811), the card would produce 210k ppd.
More to come....

Andy
 
shader +100 mhz (= 1120 mhz),
memory on stock,
Voltage = 1,18V
TDP setting = 100%
fan on 70%,
temp below 60C

Interestingly, NVidia chose 82C as default temp threshold for the GTX 780 Ti (The Titan and 780 had 80C). Anyway, with the fans set at 70%, its far away from this temp.
 
Very nice. I'm going to set my 290s up later today and see how they do.
 
DooKey said:
Very nice. I'm going to set my 290s up later today and see how they do.
Click to expand...


Nice! Anxious to see how the 290(X) both perform, although I'm hearing they're nailing 140-150K easily at stock, but with high temps.

Would kinda surprise me if GTX 780 Ti reached 210K. I know my GTX 780 hits 140K stock, the Titan around 10-20K (WU dependent) stock. I saw very little boost in PPD on my 780 bumping it up another 60Mhz (already factory OC'ed about 60Mhz), although recent drivers have since drove down the GK110 chips for some unknown reason.
 
Liger88 said:
Would kinda surprise me if GTX 780 Ti reached 210K. I know my GTX 780 hits 140K stock, the Titan around 10-20K (WU dependent) stock. I saw very little boost in PPD on my 780 bumping it up another 60Mhz (already factory OC'ed about 60Mhz), although recent drivers have since drove down the GK110 chips for some unknown reason.
Click to expand...

Which drivers? I'm on 331.65 and I find them to be better than the last couple of pre-331 drivers (331.58 is virtually identical). 331 did drive down GK104 noticeably.

I wouldn't be surprised by 210k. I have both my 780s running at 1215 MHz core. 8900 and 7811 nets about 173k, and 7810 does around 182k.
 
Zagen30 said:
Which drivers? I'm on 331.65 and I find them to be better than the last couple of pre-331 drivers (331.58 is virtually identical). 331 did drive down GK104 noticeably.

I wouldn't be surprised by 210k. I have both my 780s running at 1215 MHz core. 8900 and 7811 nets about 173k, and 7810 does around 182k.
Click to expand...


Yeah I think it was your thread on EVGA that brings the topic up. The old drivers (320.11) gave the highest PPD for both my 780 and 660, but the 36 hour bug got annoying. After that I switched to a couple others and settled on the 326.80. Immediately lost about 14 seconds on my TPF on the 780, 660 remained the same. Drivers past that seem to benefit screw up 600 series cards while boosting the GK110 core. Running multi-gpu environments there isn't yet a suitable driver it seems that maximizes both generations efficiently. Just going by the numbers of 156K stock (320.11) to 146K (326.80) a minor overclock of 60Mhz still hasn't returned my TPF and barely affected it.

Original TPF was 3:02 (P8900) and now its 3:14. That 60Mhz boost only knocked it down a second (3:13) which is why I assume driver issues are still going to be a problem. Not much for the GTX 780 Ti's that AndyE is testing though since he usually runs an entire system with similar GPU's. If the numbers hold strong it might be worth upgrading and knocking out the 660 entirely from the system to avoid this driver bug with new ones.

[EDIT]

Just for reference:

GTX 660 SC - GPU: 1011Mhz Memory: 1402Mhz (down -100Mhz for stability for some reason it'll freak out if left at default) Boost: 1149Mhz regularly while folding 24/7.
GTX 780 SC - GPU: 1002Mhz Memory: 1502Mhz Boost: 1123Mhz

I might try kicking the 780 up another 75Mhz just to see how much it helps. Just boggles me how I lost over 7 seconds from the TPF from a driver update that fixed the 36 hour bug and overclocking has so far been ineffective.
 
Last edited:
Strange observation in EVGA precision.
EVGA Precision (v4) is showing 3005 MHz memory frequency (it should be 3500 MHz)

There is a GTX Titan in the same system where I set the shader speed to +100 Mhz (=1100 Mhz) to compare the cards with identical frequency. The GTX 780 Ti is now on +80 MHz (=1100 MHz)
I will update EVGA's precision tool to the latest version to check for the 6 vs 7 GHz memory speed anomaly.

3 units done so far, with kind of high variabiility:
In the first P7810 work unit, tpf was 1:16 (210k ppd)
second P7810 WU, 1:21 and 191k ppd,
third P7810 WU 1:18 and 203k ppd respectively

to compare: The GTX Titan had 1:24 and 181k ppd

Andy
 
Andy,
Which EVGA model did you wind up with?
I believe there is more than one, but I could be thinking of the "plain" 780.
 
Core32 said:
Andy,
Which EVGA model did you wind up with?
I believe there is more than one, but I could be thinking of the "plain" 780.
Click to expand...
No EVGA model, its the Asus GTX 780 Ti - the only one available in my country for now
 
It's 2am - time for bed :)

A quick recap:
The 200k and above WUs looked like initial edge cases.
In the average the slightly overclocked Ti achieved tpfs of 1:23 (P7810) which translates into 184k ppd.
The Titan in the same system and set to the same shader frequency achieved tpfs of 1:26 and 175k ppd.
A ppd plus of 5%, a tpf plus of 3.6% (but much better price)

EVGA Precision (upgraded to the latest version) cannot properly display the memory bus frequency - its a bug.
I cross checked with GPU Shark. There the tool reports the 7 GHz memory bus speed of the Ti.

Why the Ti does not offer 7% more throughput at the same frequency vs. the Titan - I don't know yet. (15 vs. 14 SMX)

Temp in the Coolermaster BX case is well under control

original.jpg


Andy
 
Update, some numbers of WUs collected over the night

GTX 780 Ti:
P7811: 170k ppd
P7811: 175k ppd
P7810: 185k ppd
P7810: 188k ppd

GTX Titan (in the same system, shaders are at the same frequency as the GTX 780 Ti)
P7810: 181k ppd
P7810: 181k ppd
P7811: 152k ppd (?)
P8900: 175k ppd


(Unless there are some issues in my setup)
With shaders set at same frequency,and P7810 work units, the GTX 780 Ti is more or less as fast as the GTX Titan, eventually with a few percent above the Titan.
Given the cost of the card, there is progress on this front (about 30%).
In terms of energy efficiency, it seems to be less so. (Which is not bad, given the GTX Titan's excellent heritage, but no further progress beyond)

Andy
 
Very strange. What are the GPU load numbers (reported by GPU-z) for your 780Ti and for Titan when folding?
 
The system is up and running for a week now.
It is set to identical shader frequencies for the Titan and 780 Ti (1100 MHz in Boost)
No special observations to report - no system hang, etc ..

A lot of P8900 units were lately be processed.
In the average,
the GTX 780 Ti had a tpf of 2min40sec (193k ppd),
the Titan had a tpf of 2:52sec (173k ppd)

tpf is 7,5% faster
ppd is 11,6% higher (due to QRB)

With P8900 WUs, the 780 Ti seems to be approx 7% faster than the Titan based on tpf times.
Well in line with the 15 to 14 streaming multiprocessor ratio of the two GK110 GPUs (7.1%).

The GTX 780 Ti temperature is between 58C and 62C (with fan set to 70% )

Andy
 
Santa left an early present for me:

3v76.jpg


It's driving the points calculator on the v7 client bonkers. :eek:
 
Does anyone notice that 780 Ti has been released for more than two weeks, but the official linux driver for 780 Ti is not released till now? :confused:
Go to http://www.geforce.com/drivers, Manual Driver Search, select Geforce GTX 780 Ti and Linux 64-bit and submit. The result is:
No drivers found.
Please select a different combination or use the automatic driver detection option.
Click to expand...
 
quickz said:
Does anyone notice that 780 Ti has been released for more than two weeks, but the official linux driver for 780 Ti is not released till now? :confused:
Go to http://www.geforce.com/drivers, Manual Driver Search, select Geforce GTX 780 Ti and Linux 64-bit and submit. The result is:
Click to expand...

Right, even the nvidia website that is usually better updated than the geforce one does not list the 780 Ti as supported by the 331.20 driver.

Though... it appears phoronix used this very same 331.20 binary driver with Ubuntu 13.10/kernel 3.12 without any problems ?
 
Revisiting this thread, if I wanted a great folding GPU, it appears price/performance favors the vanilla 780. If I read things right, the 780 TI is 40% more expensive and delivers ~25% folding performance increase over the 780. Is that correct? Which vendor's 780 is best for overclocking? I saw Zagen30 is running at 1215MHz. Nice!
 
brilong said:
Revisiting this thread, if I wanted a great folding GPU, it appears price/performance favors the vanilla 780. If I read things right, the 780 TI is 40% more expensive and delivers ~25% folding performance increase over the 780. Is that correct? Which vendor's 780 is best for overclocking? I saw Zagen30 is running at 1215MHz. Nice!
Click to expand...



One thing I was wondering is whether or not Zagen30 was referring to stock speed w/o boost or the ultimate 24/7 boost speed.
[edit: confirmed it was based off Boost clock not stock]

I'm running the 780 SC from EVGA:

http://www.newegg.com/Product/Product.aspx?Item=N82E16814130918

I haven't really even bothered much with the OC (already well over the stock 967Mhz vs 863Mhz) and I've bumped it up to +70 (1067Mhz Core) and it Boosts to 1188Mhz (24/7 when folding). Haven't run into any issues at all. Pretty sure I could bump it up so the boost equals that 1215Mhz.

The MSI Lightning have finally come down to a more appropriate price would probably be the best of the best though. I've seen some incredible OC's on that card.

5nh.png
99d.png
 
Last edited:
ChristianVirtual said:
With or without 36hour freeze bug ? Given the smaller version number I fear the 36hour are not fixed ... :confused:
Click to expand...
What is the 36 hrs freeze bug?

My Ti is folding since 17 days without issues - not a single reboot of the system (one Ti and one Titan are in this box).

Should I expect it to freeze in the next 36 hrs?

rgds,
Andy
 
AndyE said:
What is the 36 hrs freeze bug?

My Ti is folding since 17 days without issues - not a single reboot of the system (one Ti and one Titan are in this box).

Should I expect it to freeze in the next 36 hrs?

rgds,
Andy
Click to expand...

That's interesting ? What driver version do you use ?

NV driver like 319.x and later had the issue that under Windows and Linux the GPU could only run for quite exactly 36 hours. Then the system froze, first blocking USB (no keyboard or mouse) and short after also network died leaving only the hard reset to reanimate.

For the GK110 card like the "old" 780 that is kind of solved with 331.x. On the other side those driver leave a performance hit on older cards.
With the 331.x drivers by twin 780 run for days without issues. But I had to remove my 660Ti from the same system as it dropped form 64kPPD to approx. 23kPPD. The 660Ti is back to driver 319.x on a dedicated motherboard.

See also here http://hardforum.com/showthread.php?t=1775998&highlight=
Or
https://foldingforum.org/viewtopic.php?nomobile=1&f=80&t=25052
https://foldingforum.org/viewtopic.php?nomobile=1&f=80&t=25203
 
Good information in this thread. My 780 Ti has slightly better results folding, but it's overclocked more. Seems every little bit you can overclock dramatically improves your TPF.
 
AndyE said:
With shaders set at same frequency,and P7810 work units, the GTX 780 Ti is more or less as fast as the GTX Titan, eventually with a few percent above the Titan.
Given the cost of the card, there is progress on this front (about 30%).
In terms of energy efficiency, it seems to be less so. (Which is not bad, given the GTX Titan's excellent heritage, but no further progress beyond)
Andy
Click to expand...

AndyE said:
With P8900 WUs, the 780 Ti seems to be approx 7% faster than the Titan based on tpf times.
Well in line with the 15 to 14 streaming multiprocessor ratio of the two GK110 GPUs (7.1%).
Click to expand...

Thanks for sharing results.

I can't pinpoint the performance you are getting at a 1:1 clock setup, but the Titan has one edge over the 780 TI despite a lower shader count.

Titan can process FP64 at 1/3 FP32, while the 780 TI only runs FP64 at 1/24 FP32.

I do not know the composition of calculations in P8900, but if you hit a lot of FP64 calculations the Titan will have a clear edge.
 
runs2far said:
Titan can process FP64 at 1/3 FP32, while the 780 TI only runs FP64 at 1/24 FP32.

I do not know the composition of calculations in P8900, but if you hit a lot of FP64 calculations the Titan will have a clear edge.
Click to expand...

Correct. But, ....

Folding doesn't use DP (Double Precision) calculation.
For other apps, it might be relevant.

To accomodate the "unbalanced" SP/DP ratio of many GPU cards, there is a new emerging field called mixed precision algorithms - you might check if interested.

rgds,
Andy
 
OK, thx

As I wrote I do not know what data types P8900 uses :)

If it's all int and FP32, then the 780TI should have the 7% edge over Titan, when all clock are equal.
Looks like that is what you have described :)
 
Hi AndyE,

It looks like your 780 Ti's PPD number is about 10 - 20% lower than it should be.
Maybe you should give a try to Linux, a 780 Ti at default clock is known to capable of producing as high as ~200k PPD on P8900 under Linux, with the 331.20 driver.

AndyE said:
A quick recap:
The 200k and above WUs looked like initial edge cases.
In the average the slightly overclocked Ti achieved tpfs of 1:23 (P7810) which translates into 184k ppd.
The Titan in the same system and set to the same shader frequency achieved tpfs of 1:26 and 175k ppd.
A ppd plus of 5%, a tpf plus of 3.6% (but much better price)

EVGA Precision (upgraded to the latest version) cannot properly display the memory bus frequency - its a bug.
I cross checked with GPU Shark. There the tool reports the 7 GHz memory bus speed of the Ti.

Why the Ti does not offer 7% more throughput at the same frequency vs. the Titan - I don't know yet. (15 vs. 14 SMX)
Click to expand...
 
Last edited:
quickz said:
Hi AndyE,

It looks like your 780 Ti's PPD number is about 10 - 20% lower than it should be.
Maybe you should give a try to Linux, a 780 Ti at default clock is known to capable of producing as high as ~200k PPD on P8900 under Linux, with the 331.20 driver.
Click to expand...
Hi Quickz,
thanks for the heads-up.

Due to your post I checked my systems. Currently all 10 NVidia cards are set to run at 1097 MHz, a speed which doesn't require attention. Not a single hiccup in the last 5 days since I started them for the "default case".

With that speed, the 780 Ti takes about TPF of 2:37 per P8900 unit (checked in the logs of finished WUs). This would translate to 199k ppd, about the level you are referring to.

While checking the Titan's, they seem to be lower than I would expect in relation to the Ti:
I saw TPFs of 2:50 to 2:57, translating to 166k - 176k ppd

If I would take the 7% perf difference due to shader count difference, TPF should be in the 2:47 range (181k ppd).

Andy
 
AndyE said:
Got my GTX 780 Ti today.
original.jpg


Tried this afternoon to run some folding WU, but it wasn't included in the GPUs.txt file. Thanks to the ff team for quickly expanding the file to include the card. It is now working on its first units.

You will be pleasantly surprised :)

Based on approx 20% of the first WU (P7811), the card would produce 210k ppd.
More to come....

Andy
Click to expand...

;) Nice!!
 
Current stats for my EVGA 780ti SC ACX:

Clock: 1137, which is "stock" for the SC
Driver: 331.82

Typical tpf for a 8900 unit is 2:28

According to EOC stats, the 24 hour average for my card is around 215k + ppd. That is close to half of the ppd of my 4P while it chokes on a 8101. :eek:
 
You must log in or register to reply here.
Back
Top