• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

GPU overclocking

M

mwarps

Supreme [H]ardness
Joined
Oct 6, 2002
Messages
7,072
I was just running a few basic calculations.

If a modern vid card, running around 500MHz core speed throws out 500GFLOPS, that's about 1000 flop/cycle.

Overclocking your GPU by 2 MHz would be like borging a new PC.

I think I need to change my pants.



 
Not to burst any bubbles, but I think thats a peak performance, best case scenario, zero overhead figure.. I imagine that it will be closer to 250GFLops, meaning.. you'll have to raise the core speed a whopping 4MHz to achieve the processing power of an additional PC.
 
even if it's 4MHz.. I have been scouring for OC figures with watercooling. Some are getting almost 800 MHz core compared with 650 stock. Sooooo very exciting :)
 
The scaling is not quite that simple gents as you are using very much peak rates. We do scale close to linear with core clock, memory clock is less important at the moment. The boards are already running a little warm, and we've had correctness issues ~100C, but a default 3D clocks and voltages, we stay below 90C with the stock reference cooler. We do run significantly cooler than 3DMark06 and Oblivion.
 
Not a linear increase, pooh.
Oh well. Overclocking will be done under water for me, I can't stand heat.
 
mwarps said:
Not a linear increase, pooh.
Oh well. Overclocking will be done under water for me, I can't stand heat.
Click to expand...

Just to play it safe. How about doing a few reps on the GPUBench so you won't hurt yourself. ;)
 
mwarps said:
even if it's 4MHz.. I have been scouring for OC figures with watercooling. Some are getting almost 800 MHz core compared with 650 stock. Sooooo very exciting :)
Click to expand...
I can hit 700 Core and 690 Mem on air easily... Although that's the highest I can push my Video card with the Catalyst Control Panel... Not a big fan of ATi Tools here. Aftermarket cooling of course.
 
mwarps said:
even if it's 4MHz.. I have been scouring for OC figures with watercooling. Some are getting almost 800 MHz core compared with 650 stock. Sooooo very exciting :)
Click to expand...

This is why I wish I could use my 7900GT. Volt mods FTW! 450mhz to 680mhz is none to shabby.
 
Macaholic said:
Just to play it safe. How about doing a few reps on the GPUBench so you won't hurt yourself. ;)
Click to expand...

95% of the people here have never heard of this.

How about leaving some useful instructions for us instead of a drive-by spamming? Please don't say "read the docs", either. These docs are obviously written by the developer and useless to people who are "trying not to hurt themselves" :rolleyes:

Downloaded it, ran everything... Doesn't tell me anything useful since it requires an intimate knowledge of what it's doing
 
Yes, we don't produce a single number as that is bogus for what we do. GPUBench is designed to let us understand the architectures better, and specifically how the memory system works, how the instruction issue works, and how fast readback and download are so we know the costs of offloading to the board.

But, you can play with the clocks on your board and run the instruction issue test with MAD's to see what your board can crank out in terms of FLOPs.

Or, you can run the GPUBench.pl script and generate the webpages you see in the results section on the GPUBench site.
 
mhouston said:
Yes, we don't produce a single number as that is bogus for what we do. GPUBench is designed to let us understand the architectures better, and specifically how the memory system works, how the instruction issue works, and how fast readback and download are so we know the costs of offloading to the board.

But, you can play with the clocks on your board and run the instruction issue test with MAD's to see what your board can crank out in terms of FLOPs.

Or, you can run the GPUBench.pl script and generate the webpages you see in the results section on the GPUBench site.
Click to expand...

Ah, very good. Time to get perl on my windows box again, looks like :D

EDIT: No, never mind. making the program work on win32/activeperl5.8.x should not be *this* hard.

Somewhere in the documentation provided, there should be some sort of simple "getting started" sort of thing, listing requirements (jgraph and epstopdf to say the least) such as which distribution of perl will acutally read the "/" as opposed to "\" .. I don't want to flame here, but again, a very good percentage of us are not running development environments on our folding machines, or even our main boxen.

Thankfully, all the little applets seem to work fine. Just the perl is fubar.
 
those gpubench numers are interesting. apparently for subtraction the 7800GTX is beyond excellent??
 
@mwarps: point taken, but GPUBench was really setup for GPU developers, not end users. I thought the readme said what tools were needed, but that is probably only in the CVS tree. We have't done another release since most of the GPGPU folk just grab from CVS.

@drizzt81: We fixed this just recently. Nvidia was correctly optimizing out the shader based on the constants we were passing in. We just haven't rerun the graphs.
 
mhouston said:
@mwarps: point take, but GPUBench was really setup for GPU developers, not end users. I thought the readme said what tools were needed, but that is probably only in the CVS tree. We have't done another release since most of the GPGPU folk just grab from CVS.

@drizzt81: We fixed this just recently. Nvidia was correctly optimizing out the shader based on the constants we were passing in. We just haven't rerun the graphs.
Click to expand...

Yeah, I figured that end users really don't need this sort of information.

Fiddled with some of the thingies, and my 7900GS Go is about 7.5GFLOPS.. Slow I am assuming for this sort of thing we're doing :p

I am guessing that the reason Macaholic brought the program into the thread was because of the error calculations.. Just so we could see that the GPU isn't producing wildly insane results.
 
mhouston said:
@drizzt81: We fixed this just recently. Nvidia was correctly optimizing out the shader based on the constants we were passing in. We just haven't rerun the graphs.
Click to expand...

thanks for that info. And welcome to the forums :)
 
mwarps said:
Yeah, I figured that end users really don't need this sort of information.

Fiddled with some of the thingies, and my 7900GS Go is about 7.5GFLOPS.. Slow I am assuming for this sort of thing we're doing :p

I am guessing that the reason Macaholic brought the program into the thread was because of the error calculations.. Just so we could see that the GPU isn't producing wildly insane results.
Click to expand...

That sounds too low, my guess is that's 7.5 GInst/sec, so multiply by 8 for GFlops.

The FAH client does do sanity checking on it's own, but you may not find out about the error until you've put in a bunch of work. The issue with using something like 3DMark or a game is that you might not visually notice errors unless things are REALLY bad. Since we are a numerical app, we'll notice this sooner. My guess is this will also effect GPU physics acceleration. The issue is that while we could do something like Prime95 for GPUs, basically something to test heat and numerics, it will not be an exact match for every GPGPU app. Hrmm, maybe we can throw something together.
 
mhouston said:
That sounds too low, my guess is that's 7.5 GInst/sec, so multiply by 8 for GFlops.

The FAH client does do sanity checking on it's own, but you may not find out about the error until you've put in a bunch of work. The issue with using something like 3DMark or a game is that you might not visually notice errors unless things are REALLY bad. Since we are a numerical app, we'll notice this sooner. My guess is this will also effect GPU physics acceleration. The issue is that while we could do something like Prime95 for GPUs, basically something to test heat and numerics, it will not be an exact match for every GPGPU app. Hrmm, maybe we can throw something together.
Click to expand...

It might be a good idea to put something like that together, because I know there are a lot of us here who might want to try it, we're a pretty crazy bunch.
 
Last edited by mhouston : Today at 04:09 PM. Reason: I need spell check in posts...
Click to expand...

Google toolbar has spellcheck, can't live without it. All thou, never really use it here...:)

+1 for a Prime95 for GPU's.
 
I think they might have a Linux client out for FAH, you might want to check with Stanford... :p
 
Yes, but SPIRE is a research system that we have to run performance runs on, so no folding. Since the quarter has started again, that machine will start cranking on research soon. But, we did help the folding folks build a nice GPU cluster for folding.
 
If possible, can you provide any info on the GPU cluster you guys keep refering too. We're always interested in folding porn here.

Links/pics/root access- we're flexiable.

Just giving you a hard time on SPIRE. :)
 
How about LLNL's Gauss. You can put 512 R580's in that thing. That design was a follow-on to the SPIRE design, built by Graphstream.

http://www.graphstream.com (image on front page)

Too bad we all can't have a machine like that...
 
Interesting the install mix of both ATI and Nvida systems.
 
I have a question about that GPUBench:

the last test in the set checks for precision of the output. Would it make sense to take a close look at this when overclocking? If precision drops -> clockspeed to high?
 
Not really. That is basicaly a test for how close to IEEE the GPU is. As people overclock, the hope is that the tests built into the FAH client will catch major issues. However, it may take quite awhile for things to go wrong. But, usually when things go bad, they go REALLY bad.
 
mhouston said:
Not really. That is basicaly a test for how close to IEEE the GPU is. As people overclock, the hope is that the tests built into the FAH client will catch major issues. However, it may take quite awhile for things to go wrong. But, usually when things go bad, they go REALLY bad.
Click to expand...
Since it obviously is detrimental to the F@H team to submit WRONG results, rather than none, is there any recommendation on how to test for 'long-term' stability? In the end, I "need" to overclock my GPU (why else did I pay for that watercooling loop) yet would love to fold... or should I just wait for "gpu prime95" to be released at some point in the future and have per-game custom settings?
 
drizzt81 said:
Since it obviously is detrimental to the F@H team to submit WRONG results, rather than none, is there any recommendation on how to test for 'long-term' stability? In the end, I "need" to overclock my GPU (why else did I pay for that watercooling loop) yet would love to fold... or should I just wait for "gpu prime95" to be released at some point in the future and have per-game custom settings?
Click to expand...

Pretty much a similar response here. :D
I have a need to overclock based on a watercooling system as well.

I think something that we can do in the meantime is run a bunch of the "precision" benches from GPUbench and run a diff on the outputs from them. If they change, the OC is bad. Please note this is just an idea and I don't know precisely how the "precision" program works :)
 
mwarps said:
Pretty much a similar response here. :D
I have a need to overclock based on a watercooling system as well.

I think something that we can do in the meantime is run a bunch of the "precision" benches from GPUbench and run a diff on the outputs from them. If they change, the OC is bad. Please note this is just an idea and I don't know precisely how the "precision" program works :)
Click to expand...
that is exactly what I asked here:
drizzt81 said:
I have a question about that GPUBench:

the last test in the set checks for precision of the output. Would it make sense to take a close look at this when overclocking? If precision drops -> clockspeed to high?
Click to expand...

and got this response to it, which says "it doesn't work like that":
mhouston said:
Not really. That is basicaly a test for how close to IEEE the GPU is. As people overclock, the hope is that the tests built into the FAH client will catch major issues. However, it may take quite awhile for things to go wrong. But, usually when things go bad, they go REALLY bad.
Click to expand...

Maybe the precision benchmarks only runs a few iterations, instead of a 'stress test' ?
 
Correct, the GPUBench stuff won't heat up a board enough to be interesting. I'll see if we can come up with something numerically sensitive that we can run deep enough to check for badness. This may take awhile as getting the FAH client stable is primary right now, that and my thesis. ;-)
 
mhouston said:
*snip* This may take awhile as getting the FAH client stable is primary right now, that and my thesis. ;-)
Click to expand...


*jaw agape*

Dude. You're a monster.

:D
 
When overclocking with FAH on my 1900XT, I have just been careful to adjust the fan speed so that temps stay below 75C. It REALLY helps that it has been getting down into the 30’s at night here.

Anyway, I also did some tests while overclocking, and it appears that increasing just the GPU results in about a 93% increase in performance (I.E. 10% increase in core clock speed results in 9.3% folding performance increase). That is really nice scalability. Is that what other folks have been getting? What about you at Stanford? Just thought I would share my results. :D
 
The temperature target in testing is <90C, so you can save your ears a little and let the board run hotter if you wish. If you don't mind the noise, running the chip cooler for this and games will make the chip last longer. But, we are probably talking multi-year happiness at high temps already. The main issue with running the fan high 24/7 is we don't have stats on the bearing life in the fans they use. I know from cluster experience that we are probably also talking in terms of a few years, but these boards haven't been out that long to know. ;-)

We are trying to throw together a stress test so people can see when they start throwing bad data when overclocking or running warm.
 
You must log in or register to reply here.
Back
Top