Author |
Message |
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
This is a bit of a hardware question, but I guess its on the right place here. If not a moderator may move it.
I built me a new rig last week and seeing a few things.
I had a 770 in an AMD fx8350 on a Sabertooth MOBO.
My new one has an i7 4471 on a Sabertooth MOBO as well (not the same type off course). And with the same WU's (if they are really the same) the 770 is around 3000 seconds faster on the Intel.
Secondly I put a GTX780Ti in as the primary GPU and this one is around 3000 seconds faster than the 770.
I installed all hardware and software from the box, did no tweaking and let it run. Checking with GPU-Z I saw that the 780Ti went quickly to 85-87°C, what I don't like so I used GPUTweak from Asus to make a change. I have set the GPU Temp Target to 75°C and left the Power Target to 100%. Did nothing with clock speeds. Set fan to user define and made that it run more. Now I see:
The 780Ti:
GPU Core CLock:966.6MHz
GPU Memory Clock: 1752.8MHz
Temperature: 75°C
GPU Load: 80-81%
Power Consumption: ~74%
PerfCap Reason: Limited by temperature limit (I understand as I set that)
The 770:
GPU Core Clock: 1045.2MHZ
GPU Memory Clock: 1752.8MHz
Temperature: 63°C
GPU Load: 94%
Power Consumption: ~69%
PerfCap Reason: Limited by reliability voltage (I don't understand)
The 780Ti is from Asus and has 1 fan, the 770 is also from Asus and has two fans.
When I set the temperature limit to 85°C, then the GPU load doesn't change but the card operates at 82-83°C.
Question: how can I get the 780Ti to operate around 90%? Or is ~80% its maximum with Santi WU's?
____________
Greetings from TJ |
|
|
DagorathSend message
Joined: 16 Mar 11 Posts: 509 Credit: 179,005,236 RAC: 0 Level
Scientific publications
|
I think the clue is here:
PerfCap Reason: Limited by temperature limit (I understand as I set that)
If you get the temperature down to 70C or less the card will automatically boost the GPU Core clock rate?
____________
BOINC <<--- credit whores, pedants, alien hunters |
|
|
|
"PerfCap Reason: Limited by reliability voltage"
Means the card has reached it's maximum allowed voltage for reliable operation, i.e. the top turbo boost bin, and won't go any higher. BTW: on GTX770 cards this is ~1.20 V, more than previous cards were allowed. At this voltage the card probably takes ~1.20 GHz instead of those "wimpy" 1.05 GHz. For comparison: my GK104 (same chip) reaches those 1.05 GHz at ~1.05 V.. which is incidently what I'm running it at. Far more efficent than 1.20 V.
Regarding your 780Ti: as Dagorath said if you increase the fan speed even more and lower the temperature, the card will boost higher (and use a less efficient voltage to get there, of course). However, this will not increase GPU utilization, as it's already in % and always refers to the current clock speed, not some fantasy value. I.e. if the GPU is bottlenecked somewhere GPU utilization will even drop at higher clock speeds, because there simply won'T be enough work to do for the cores.
I don't know what's holding your card back. You could stop any other work te CPU might be doing and test if this help significantly. You could scale the power limit back and watch the results: at lower clock speeds GPU utilization should increase, while efficiency will go up (until you reach the base clock & voltage). You could overclock your GPU memory and see if this helps. The cards are said to take almost 2000 MHz, but overclocking is an obvious risk. You could also try to run 2 concurrent tasks - maybe this GPu is finally large enough to benefit from this with more than SRs. Although this might change as soon as you get some Nathan-kicks.
MrS
____________
Scanning for our furry friends since Jan 2002 |
|
|
|
...The cards are said to take almost 2000 MHz, but overclocking is an obvious risk.
What cards are you referring to? 100% overclock is a bit steep.
|
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
The GPU Memory clock is 7012MHz. This is (was) standard I did not touch that setting at all. Is that okay?
I have now set the fan higher and the GPU Temp Target to 69°C, to see it that works. Currently still at 81% GPU Usage.
Tried already to stop all CPU tasks, but didn't help. Now only 4 WU's (Docking) are running on the CPU.
The PC is now running for 3 days and 7 hours. The graphics of Asus GPUTweak are no gone, empty, only the grid. But the values at the right side change. Is this normal of a bug from Asus?
Let it run overnight and look in the morning again.
Thanks so far.
____________
Greetings from TJ |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
Letting the 780Ti run overnight at 69°C resulted in an GPU load of 82-83% but the GPU Core Clock went to 885MHz.
I have now set maximum temperature to 70°C and GPU Core Clock is running at 901.4MHz, with GPU Load of 81% steady. Perhaps this 81-83% is the maximum for this card for Santi's as ETA suggested.
The VDDC is 1.0250V for the 780Ti and 1.160V for the 770 according to GPU-Z. Do I need to change anything there?
____________
Greetings from TJ |
|
|
|
Something is hindering the performance of your GTX 780 Ti. I guess that it's the Win7 WDDM overhead.
For comparison:
I163-SANTI_bax2-6-32-RND6602_0 24460s (6h48m) (your Win7 host)
I203-SANTI_bax2-4-32-RND8709_0 16905s (4h42m) (my WinXP host)
These are not some specially selected running times, I have similar running times for this kind of WU on my host, and you have similar running times on yours (so I'm not comparing my best running time to your worst.) According to these numbers, my GTX 780 Ti is 44,7% faster under WinXP (6 SIMAP CPU tasks are running) than yours under Win7. (our CPUs are almost the same, as I have an i7-4770k@3.5GHz)
My GTX 780 Ti is running at 1110 MHz, temp: 65°C, GPU load: 96%, VDDC: 1.15-1.20V, Memory controller load: 26%, (while crunching 167x-SANTI_MAR423cap310-1-32-RND5480_1) |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
Thanks for your replay Zoltan.
Indeed I have checked my times with other 780Ti's and that is why I started this thread. My 770 is almost as fast as the 780Ti and that can't be.
You have you card way cooler running then I. Can you please advise me what I need to change? What program are you using to alter the settings?
What is the Win7 WDDM overhead? And if, how can I change that?
Your replay is highly appreciated as always.
____________
Greetings from TJ |
|
|
|
You have you card way cooler running then I. Can you please advise me what I need to change? What program are you using to alter the settings?
I'm using MSI Afterburner, with user defined fan control (30°C 40% fan speed to 80°C 100% fan speed) the fan speed is 85% at the moment (and it's noisy a bit), but this configuration is caseless :)
What is the Win7 WDDM overhead?
This is the Windows Display Driver Model introduced with Windows Vista. It's a more secure way to access the display's hardware (a display hardware/software failure won't bring the whole OS down), but at the same time it's slower - hence the overhead. (as far as I understand it)
And if, how can I change that?
You can't turn it off. You can change it only by using a different OS (Linux, WinXP). That's why I'm still using WinXPx64. It will be quite exciting after April 2014, since WinXP's extended support period will expire then - so there won't be more hotfixes for WinXP (and Win.Server 2003) from MS after that. |
|
|
|
I think this WDDM overhead could impact faster cards more than slower cards, as a faster card will complete a piece of calculation faster, but the WDDM overhead of every interaction with the OS takes the same time either for a slower or faster GPU, and the number of interactions is (almost) the same for every GPU (slower or faster) while crunching a given WU. So for example the total time of the calculations on the GPU is X, and the WDDM overhead is Y, then the total running time is X+Y. If you use a twice as fast GPU, the total time of the calculations will be X/2, but the WDDM overhead stays Y, so the total running time will be X/2+Y. If the X/Y ratio is high, the total running time won't be affected by Y too much, so the running time could be reduced (almost) in direct proportion by using a faster GPU, but I'm afraid, that the GTX 780 Ti (and the GTX 770) is too fast, so the X/Y ratio is not high enough to have linear decrease in the running times (on Win7) by using a faster card. The Maxwell chips will have their own dedicated ARM CPU on chip to eliminate this bottleneck.
|
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
At the moment I have a fan speed of 82% and 70°C with Asus GPU Tweak.
When using MSI Afterburner I can not find my second card to alter things.
If I go for default setting, the GPU Boost Clock displays 982 MHZ but operates at 876 MHZ, the temperature goes to 82°C and fan duty to 60%.
GPU Load stays at 81%. So the default is not helping me.
Do you have a second card in this system Zoltan?
As I have the 780Ti in the first slot and the 770 in the second. There is about 2 centimeters room in between. The 770 (with two fans, stays cool at 61°C.
____________
Greetings from TJ |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
Just a notice.
I was my cards a bit tweaking, core clock little higher to 1045MHz (not nown if related) but all of a sudden screens went black and system boots,, with message Windows didn't close correctly.
Checked with WhoCrashed and it was a Bad Pool Header from nVidia Windows kernel driver 331.82.
After boot and everything running again, one Noelia (was 98% before crash) finished correctly, the other a Santi was 54% done, not as it started new again.
____________
Greetings from TJ |
|
|
|
Do you have a second card in this system Zoltan?
Yes I have a GTX 670 in the farthest PCIe slot (it runs at PICe2.0 x4, but the GTX 780 Ti runs at PCIe3.0 x16).
As I have the 780Ti in the first slot and the 770 in the second. There is about 2 centimeters room in between. The 770 (with two fans, stays cool at 61°C.
There is 4 slot space between the back of the GTX 670 and the "glass" of the GTX 780 Ti's cooler. It's about 9 centimeters. |
|
|
|
Just a notice.
I was my cards a bit tweaking, core clock little higher to 1045MHz (not nown if related) but all of a sudden screens went black and system boots,, with message Windows didn't close correctly.
Checked with WhoCrashed and it was a Bad Pool Header from nVidia Windows kernel driver 331.82.
After boot and everything running again, one Noelia (was 98% before crash) finished correctly, the other a Santi was 54% done, not as it started new again.
So much for the WDDM is crash free. |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
I have nothing changed at the 770, but set the voltage of the 780Ti to 1.175mV and the Boost Clock to 1110MHz. The Power Target still at 100% and the Temp Target to 70°C and let it run overnight.
According to GPU-Z the Core Clock is running at 875.7MHz, the GPU load is ~78% and Power Consumption is 57.8%. The temperature is 67°C (which I like). The PerfCap Reason is now GPU utilization.
It does now LR's in less than 7 hours. So it made a good contribution to the project.
Could it be better, if I give it little more voltage and set the core clock lower or have I already screwed up the settings?
GPU-Z has more sensors for the 770, 4 about voltage, and 1 of VRM Temperature.
I don't see this with the 780Ti and the 660. Is the 770 special?
____________
Greetings from TJ |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
There is 4 slot space between the back of the GTX 670 and the "glass" of the GTX 780 Ti's cooler. It's about 9 centimeters.
That is good, but you have another MOBO than I have, and I wouldn't listen to you :)
Well I am already saving for a Maxwell and will put that in a new system with a Gigabyte MOBO.
____________
Greetings from TJ |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
I have made the settings of the 780Ti the same as for the 770 but the clock stays operating at 875MHz.
Finally I got a Noelia on the 780Ti and now the GPU load is 91%, while on the Santi's it is ~80% but also with the Noelia the clock stays 875MHz.
Mostly Zoltan has good advise but as he runs XP, he has a more capable OS then I with Win7. Perhaps someone with Win7 and a 780Ti can help me or give some more info to steer the 780Ti to a higher GPU load on other tasks than Noelia's or perhaps it is not possible with a 780Ti and is the "to fast" for Windows?
____________
Greetings from TJ |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
Hello Zoltan,
I have found how I get my cards BIOS into the Kepler BIOS Tweaker.
If you have time (as your own problem is sort out) would you please:
1. let me know which values you have changed?
2. how then get the new cards BIOS into the card?
Thanks for you help as always.
____________
Greetings from TJ |
|
|
|
Hi TJ!
I have found how I get my cards BIOS into the Kepler BIOS Tweaker.
If you have time (as your own problem is sort out) would you please:
1. let me know which values you have changed?
I've changed everything :)
Tell me what do you want to achieve, and I'll try to be more specific :).
There are non documented settings for example some fields on the "boost states" and on the "clock states" tab.
If you want to increase the stability of your card, you should try to increase the voltage by 12mV steps, and/or the power target. The percentages shown will change beside the power values, as you change them. The power values are in miliwatts, (it's quite pointless for the high-end cards) so you can forget the last thre zeroes, as you can change it only by 1000mW (=1W) steps.
Or you can lower the boost clock ("Common" tab) and on the P0 and P2 states (Boost states tab).
I suggest to change only one stetting at a time, and keep the original BIOS.
2. how then get the new cards BIOS into the card?
1. For practical reasons, you should put nvflash.exe (plus nvflsh32.sys and nvflsh64.sys) to the same folder where the (modded) bios is located.
2. Open an administrative command prompt.
3. Go to the folder where the (modded) BIOS file is stored. (for exapmple: cd \\keplerbiostewaker)
4. If you have more than one card in the system, you should check which one is to flash with the command: nvflash -a
5. Use nvflash.exe to flash the BIOS to the card: nvflash mybios.rom
6. If you have more than one card in the system, you should add an "--index=n" switch to the command line, specifying the number of the card you got in step 4. nvflash --index=1 mybios.rom (you have to check hex numbers of vendor ID, product ID, subvendor ID, PCI subsystem ID, or PCI bus number, so it's much easier if you have only one card in the system)
7. If the BIOS update is successful, you should restart your system. |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
Thank you Zoltan.
I have changed all settings too with MSI Afterburner, GPU Tweak (from Asus) and EVGA Precision X.
I thought the 780Ti would be a lot faster then a GTX 770, but it is only ~3000 seconds faster, on Win7.
Your card is about 7000 seconds faster then mine due to XP.
What I would like to achieve is to be that the 780Ti is a bit faster. I will try Linux but don't have not a lot of confidence yet.
If you have some tips/idea's I could try. I would be grateful.
____________
Greetings from TJ |
|
|
|
I thought the 780Ti would be a lot faster then a GTX 770, but it is only ~3000 seconds faster, on Win7.
Your card is about 7000 seconds faster then mine due to XP.
You know the source of the difference: it's the operating system, not the card. There's no such tweaking of the card which could compensate the loss caused by the OS. Also, there's no way for the user to change the way the OS access the hardware it's running on, because it is the job (the main purpose and the characteristic) of the OS. The users wanted to have less BSODs, and system restarts, they got it with WDDM, but with that comes its overhead, which is bad for us, especially for the recent SANTI workunits, which interact with the CPU more than the NATHAN_KIDKIX workunits.
What I would like to achieve is to be that the 780Ti is a bit faster.
I'm sorry to say that, but you have to change the OS.
I will try Linux but don't have not a lot of confidence yet.
This is a good idea, but I can't help you with that, as I don't use Linux, and I don't have any experience with Linux. But there are a lot of other crunchers who could (and will) help you.
If you have some tips/idea's I could try. I would be grateful.
If this is a dedicated cruncher PC, you can have WinXPx64 on that.
But...
Maxwell GPUs will have a dedicated (ARM) CPU on chip, thus the WDDM could be eliminated - so those GPUs will be much faster under Win7 than the Keplers. We don't know their release date and their specifications yet, and we don't know how long it would take to make the ACEMD client to work with them, but Maxwell's architecture could be the solution for your problem. Until then, you should use a different OS. |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
Hi Zoltan,
You are right again :)
Indeed I have tried all I could and got the core clock steady at 980MHz, even managed to get it to 1045, but the GPU load is around 74% with Santi's, and 10% higher with Noleia's and Nathan's.
I will start experimenting with Linux first, and wait with investment in new cards until that works smooth or the Maxwell's arrive.
Your advice was of great help.
____________
Greetings from TJ |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
With the Noelia FXArep WU's my 780Ti is approximately 5000 seconds faster than a Santi WU on Windows7!
So it is not only the WDDM that slows down the card of cards it is also a bit of programming. Noeilia does that great. The 780Ti runs at 82% reaches 85% GPU load at times (Santi only 77%) and at 72-74°C.
With Win7 I can use the iGPU for Einstein with Linux an iGPU is useless at the time (if I have read correct).
____________
Greetings from TJ |
|
|
skgivenVolunteer moderator Volunteer tester
Send message
Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level
Scientific publications
|
You need to compare like for like Work Units on each operating system. At present my W7 system is using 81% and 86% of the GPU's while running a Santi and Noelia WU respectively, on a GTX770 and a GTX670. Thus Noelia's present WU's better utilize the GPU by 6%. The runtimes will vary for many different reasons (size, complexity, PCIE dependency...). Interestingly Noelia's WU's at one time struggled to get anywhere near 70% GPU utilization, and didn't use the CPU fully! Now they do use a 'full' CPU core/thread and do utilize much better.
Last time I checked like for like WU's on an identical system and W7 vs Linux, I found that a GTX770 was 12.5% faster on Linux. I expect it would be slightly different for different computer architectures (hardware; CPU, Chipset...), but I expect that this may not quite be linier; so the better the card the larger the gap. So the OS difference for a GTX650 might just be 10% while for a GTX680Ti it could be 15%. This would really need to be tested on different hardware, and a 5% difference might be hard to spot, so trying to work this out with just 670's and 770's would just be guesswork.
____________
FAQ's
HOW TO:
- Opt out of Beta Tests
- Ask for Help |
|
|
Jozef JSend message
Joined: 7 Jun 12 Posts: 112 Credit: 1,140,895,172 RAC: 17,863 Level
Scientific publications
|
Hi I have this graphics card running on Windows 8.1 64 bit, Gigabyte GeForce GTX 780 Ti WindForce 3X OC.
All tasks short or long in Seconds after the initiation of counting ,they fall .. I have the latest drivers,update, etc. ..
The same problem I have also another type of BOINC projects on 55 cuda. Other boincprojects going on 50 cuda working normally.
interesting is the second PC where there are 3 Nvidia GTX 680, the basic version, not OC versi. These work normally while. also the same windows and drivers.
What to do to give me my 780 Ti start counting tasks.? |
|
|
|
Hi I have this graphics card running on Windows 8.1 64 bit, Gigabyte GeForce GTX 780 Ti WindForce 3X OC.
All tasks short or long in Seconds after the initiation of counting ,they fall .. I have the latest drivers,update, etc. ..
The same problem I have also another type of BOINC projects on 55 cuda. Other boincprojects going on 50 cuda working normally.
interesting is the second PC where there are 3 Nvidia GTX 680, the basic version, not OC versi. These work normally while. also the same windows and drivers.
What to do to give me my 780 Ti start counting tasks.?
I had the same the problem with my same type of card.
I've started a thread about it, but the solution is that:
You have to lower the card's memory clock to 2700MHz. You can do it by the following steps:
1. read the BIOS with GPU-Z
2. modify the BIOS with Kepler BIOS tweaker 1.26 (you have to log in to download the zip file at the end of the post)
3. flash the BIOS with nvflash |
|
|
|
Or just edit the GPU memory clock in software and see if it helps. If it does, you can set this automatically upon boot. Well.. actually, if the GPU memory can't function error-free at specified frequency that would be a reason for an RMA. And if the next card also can't do it maybe nVidia pushed memory clocks too high.
@TJ: on POEM, folding@home and possibly other projects the complete 331 / 332 driver branch is significantly slower than the previous version, we're talking about ~15% here. That's why I'm "still" using 327.23. Not sure if this one already supports GK110 on GPU-Grid. If so you might want to give this one a try. I haven't read anything about similar driver problems here, but I have been busy otherwise the last months or so.
MrS
____________
Scanning for our furry friends since Jan 2002 |
|
|
TJSend message
Joined: 26 Jun 09 Posts: 815 Credit: 1,470,385,294 RAC: 0 Level
Scientific publications
|
@TJ: on POEM, folding@home and possibly other projects the complete 331 / 332 driver branch is significantly slower than the previous version, we're talking about ~15% here. That's why I'm "still" using 327.23. Not sure if this one already supports GK110 on GPU-Grid. If so you might want to give this one a try. I haven't read anything about similar driver problems here, but I have been busy otherwise the last months or so.
MrS
Thanks for your input ETA. Indeed I missed your suggestions/comments for a while.
I did try it with lower drivers, but then the 780Ti didn't work. Not sure if it was the 327.23 so I will try that one next time the rig needs booting due to Windows updates.
____________
Greetings from TJ |
|
|