GPU block testing

pHaestus · 10-24-2004, 02:36 PM

In principle testing a GPU wb isn't so different from testing a CPU wb. You still need to measure temperatures at the same spots, and you still need to measure dP and vary flow rates to actually characterize performance.

But what of the proper die size/heat source? It's possible to test the blocks on a real video card (I have a GF3Ti200 that outputs a diode reading, 9800 pro and newer ATI cards do this for certain too). There's no way to calibrate this reading though. I could use a die simulator designed originally for SocketA CPUs, but then the heat density is much much different from the real world system.

One rather unorthodox idea would be to use the cold plate of an MC50-T GPU wb as a heat source by reversing the TEC so that it's chilling water of one loop and being cooled by the GPU wb I'm testing. This isn't an ideal heat source but frankly it seems like my best bet atm (I could measure air and water temps on both sides of loop, V and A on TEC, and flow on the test loop side). Better performing wbs would produce cooler water on the cold side. The disadvantage is this is complicated.

I could also rig up a mounting method and just mount the GPU wbs on my test system CPU I guess. Again the die size is far removed from a GPU.

How would you recommend doing this? I will buy a new video card if I need to, but Christ the 6800 or X800 pro are CAD$500 and up!

jaydee · 10-24-2004, 04:48 PM

Man another monumental task your geting into.

Why is there no way to calibrate the cards diode? I assume there is no way to hack it to an external diode reader like you do with the CPU?

pHaestus · 10-24-2004, 05:17 PM

yea that's exactly the problem. Hmm I wonder if maybe some of the video card mfgrs could be of help? I'll contact a few to see...

MaxxxRacer · 10-24-2004, 08:49 PM

they would give u that info?? prolly only if u agreed to nda...

MaxxxRacer · 10-24-2004, 09:08 PM

oh, i mentioned making hte core larger on the socket a die simulator. is it possible to do this. maybe model something close to the x800 and 6800. a happy medium between the two of them. im sure someone here can provide specs on either one. Do that and have the die simulator draw 70 watts, which is what the x800 draws. I say what it draws because no one really knows exactly how much it puts out of what it draws as heat. its not like its a prescott and we know that it dumps like 90 percent of what it draws as heat... heheh.

DryFire · 10-24-2004, 09:31 PM

how exactly would you test blocks like the nv68 from DD block that also cool the ram and the extra heat could change how well it cools the core... well i guess it wouldn't make too much of a difference.

why couldn't you test all teh blocks on the socket a heat die. that way, at least, they would all use the same heatsource. Arn't different cores different sizes anyway?

MaxxxRacer · 10-24-2004, 10:24 PM

well to be honest, i dont think that the nv86 block was going to be considered. all of the blocks currently that we are going to test will be non specific in the card they cool. IE they can be used for nvidia or ati. a seperate article would have to be used for the nv68 in which you actually did use the card to benchmark it.

You could do so by looking at the temp reading the card gives off, and how much you can oc the card with either cooler. The either being the swifty block, as it is the only other block that has mounting holes for it. you can put some other water cooelrs on there, but u have to zip tie them on.. the guys over at xbit labs did it.

Cathar · 10-24-2004, 10:47 PM

Modern CPU dies are up around the 250mm^2 mark. The X800XT is 257mm^2, while the 6800 GPU is 287mm^2.

This is pushing the outer limits of profitability for silicon manufacture as things start to get pretty expensive in terms of die yield as you get much above 250mm^2.

GPU dies almost always seem to be square as well.

Would suggest perhaps an artifical heat die of 16x16mm in size. This would be a useful test tool, not just for GPU blocks, but also for CPU blocks where the present testbed size of 88mm^2 is pretty small in comparison to the ~170-220mm^2 CPU dies that are presently floating about the marketplace.

Would be good to have a "small die" and "large die" set of tests, and test all blocks on both to provide a fair idea of the relative strengths in each scenario.

Who knows - could find out that a few GPU blocks actually make for some pretty decent small CPU die blocks too.

pHaestus · 10-24-2004, 10:51 PM

Quote:

You could do so by looking at the temp reading the card gives off, and how much you can oc the card with either cooler.

Reliability of either method is potentially pretty poor. Still contemplating this but leaning towards some sort of simulator. What I have here is a JoeC style AMD die:

9mmCensor · 10-25-2004, 12:05 AM

pHaestus maybe contact ATI. They are local (to Canada). Maybe they could sponsor the GPU Cooler testing, you test on their cards, with waterblocks. And because they aren't making waterblocks their is no bais involved. And they get PR.

MaxxxRacer · 10-25-2004, 02:59 AM

I doubt ati would get anywhere near us. They dont want people frying their cards with water. And this would have them supporting water cooilng their cards. Cool idea though. Would be nice to get pH a free x800xt for his efforts.

pH, cathar uses the max OC method to test his cpu blocks. That is why I mentioned it. Between the G4 and G5 he got like 90mhz increase or something... Cant remember exactly what the increase was.. Anyway, it isnt exactly scientific, and to remain credible as you are its best to stick with numbers.

As I aksed before, is there anyway to mod ur die simulator to output the right heat and make it approximatly between the die size of a x800 and 6800. right inbetween those. would be better than a 9800 core as it would make the review valid for a longer period of time.

Cathar · 10-25-2004, 03:14 AM

Quote:

Originally Posted by MaxxxRacer

Between the G4 and G5 he got like 90mhz increase or something...

Offtopic - but I haven't been able test peak G5 overclock yet. Got the wrong motherboard and CPU installed for that at present.

MaxxxRacer · 10-25-2004, 03:27 AM

hrm... well you posted something about getting a higher oc between 2 versions of the storm. you posted on the ausie forums with some pics of cpuz if im not mistaken. correct me if im wrong. This was a while ago btw..

Offtopic again - cathar, thanks for making the storm, you are singlhandedly powering the foward design of waterblocks. (gotta throw some credit over to swifty as they are one of the few desings that didnt rip off cathar)

Cathar · 10-25-2004, 04:02 AM

Quote:

Originally Posted by MaxxxRacer

hrm... well you posted something about getting a higher oc between 2 versions of the storm. you posted on the ausie forums with some pics of cpuz if im not mistaken. correct me if im wrong. This was a while ago btw..

May be referring to Cascade vs Storm/G4. Have only had the proper Storm/G5 in working order for about 5 days now.

Back on topic - overclocking is a useful tool to analyse the performance of a cooling effect, however it is also somewhat random. Instability when overclocking is somewhat of a Poisson distribution as you get close to the edge. There's always a chance that it's unstable, but you may not see it inside of 24 hours, but the next time you run the test in might happen within 5 minutes. Would need to hold the water temperature exactly steady for that long. I like to use overclocking though as a supplementary tool to temperatures.

For GPU's it's a little trickier though. Fairly easy to catch the result of a wrong calculation (eg. SuperPI or Prime95 Torture Test) but hard to visually gauge borderline instability with video cards when the difference may be a couple of pixels per frame. There are some auto-overclocking utilities for video cards that do perform stability tests and automatically find the peak stable overclock by analysing rendered 3D frames against a reference image. That may be a useful tool to add into the mix if using real video cards. The problem with using real video cards though is that they are arguably obsoleted every 6 months or so.

pHaestus · 10-25-2004, 02:16 PM

Yes I am familiar with the ATITOOL for ATI cards. Is there a similar app for NVIDIA cards?

Blackeagle · 10-31-2004, 02:54 PM

pH,

If you go forward with GPU block testing, and I hope you do, could you run the Fusion as one of the early blocks? It's design is one of the more advanced, and it' populer so would give a good start to a graph much like the one in your CPU block articles.

I'm sure there are others you wish to do such as the Maze 4 & Swifty designs which would also be good for comparing. I just happen to like the fact that the Fusion has a brass top, so no corrosion risk. And it's the only GPU block I know that offers that, though I may have missed others.

mastermind2004 · 10-31-2004, 03:26 PM

Quote:

Originally Posted by pHaestus

Yes I am familiar with the ATITOOL for ATI cards. Is there a similar app for NVIDIA cards?

I've used Coolbits, which is just a registry hack for Nvidia cards. I believe Powerstrip works as well. If, that is, ATITOOL is just an overclocking app.

W1zzard · 01-22-2005, 04:40 PM

if you need any customized atitool versions for your testing let me know

black_dante · 01-22-2005, 05:48 PM

Quote:

I doubt ati would get anywhere near us. They dont want people frying their cards with water. And this would have them supporting water cooilng their cards. Cool idea though. Would be nice to get pH a free x800xt for his efforts.

ATI does like water and overclocking.... they use swiftech in their rigs...

MaxxxRacer · 03-14-2005, 03:24 AM

pH, I am going to contact ATI on your behalf if you dont mind. Also on my behalf i suppose as I will be doing gpu block testing with an x800xt soon enough.

10-24-2004, 02:36 PM	#1
pHaestus Big Player Making Big Money Join Date: Aug 2001 Location: irc.lostgeek.com #procooling.com Posts: 4,782	GPU block testing In principle testing a GPU wb isn't so different from testing a CPU wb. You still need to measure temperatures at the same spots, and you still need to measure dP and vary flow rates to actually characterize performance. But what of the proper die size/heat source? It's possible to test the blocks on a real video card (I have a GF3Ti200 that outputs a diode reading, 9800 pro and newer ATI cards do this for certain too). There's no way to calibrate this reading though. I could use a die simulator designed originally for SocketA CPUs, but then the heat density is much much different from the real world system. One rather unorthodox idea would be to use the cold plate of an MC50-T GPU wb as a heat source by reversing the TEC so that it's chilling water of one loop and being cooled by the GPU wb I'm testing. This isn't an ideal heat source but frankly it seems like my best bet atm (I could measure air and water temps on both sides of loop, V and A on TEC, and flow on the test loop side). Better performing wbs would produce cooler water on the cold side. The disadvantage is this is complicated. I could also rig up a mounting method and just mount the GPU wbs on my test system CPU I guess. Again the die size is far removed from a GPU. How would you recommend doing this? I will buy a new video card if I need to, but Christ the 6800 or X800 pro are CAD$500 and up!

10-24-2004, 04:48 PM	#2
jaydee Put up or Shut Up Join Date: Dec 2001 Location: Spokane WA Posts: 6,506	Man another monumental task your geting into. Why is there no way to calibrate the cards diode? I assume there is no way to hack it to an external diode reader like you do with the CPU? Last edited by jaydee116; 10-24-2004 at 06:24 PM.

10-24-2004, 08:49 PM	#4
MaxxxRacer Cooling Savant Join Date: Sep 2004 Location: Los Angeles, Ca Posts: 310	they would give u that info?? prolly only if u agreed to nda... __________________ Excuse me, I believe you have my stapler.

10-24-2004, 09:08 PM	#5
MaxxxRacer Cooling Savant Join Date: Sep 2004 Location: Los Angeles, Ca Posts: 310	oh, i mentioned making hte core larger on the socket a die simulator. is it possible to do this. maybe model something close to the x800 and 6800. a happy medium between the two of them. im sure someone here can provide specs on either one. Do that and have the die simulator draw 70 watts, which is what the x800 draws. I say what it draws because no one really knows exactly how much it puts out of what it draws as heat. its not like its a prescott and we know that it dumps like 90 percent of what it draws as heat... heheh. __________________ Excuse me, I believe you have my stapler.

10-24-2004, 10:24 PM	#7
MaxxxRacer Cooling Savant Join Date: Sep 2004 Location: Los Angeles, Ca Posts: 310	well to be honest, i dont think that the nv86 block was going to be considered. all of the blocks currently that we are going to test will be non specific in the card they cool. IE they can be used for nvidia or ati. a seperate article would have to be used for the nv68 in which you actually did use the card to benchmark it. You could do so by looking at the temp reading the card gives off, and how much you can oc the card with either cooler. The either being the swifty block, as it is the only other block that has mounting holes for it. you can put some other water cooelrs on there, but u have to zip tie them on.. the guys over at xbit labs did it. __________________ Excuse me, I believe you have my stapler.

10-24-2004, 05:17 PM	#3
pHaestus Big Player Making Big Money Join Date: Aug 2001 Location: irc.lostgeek.com #procooling.com Posts: 4,782	yea that's exactly the problem. Hmm I wonder if maybe some of the video card mfgrs could be of help? I'll contact a few to see...

10-24-2004, 09:31 PM	#6
DryFire Cooling Savant Join Date: Sep 2004 Location: USA Posts: 148	how exactly would you test blocks like the nv68 from DD block that also cool the ram and the extra heat could change how well it cools the core... well i guess it wouldn't make too much of a difference. why couldn't you test all teh blocks on the socket a heat die. that way, at least, they would all use the same heatsource. Arn't different cores different sizes anyway?

10-24-2004, 10:47 PM	#8
Cathar Thermophile Join Date: Sep 2002 Location: Melbourne, Australia Posts: 2,538	Modern CPU dies are up around the 250mm^2 mark. The X800XT is 257mm^2, while the 6800 GPU is 287mm^2. This is pushing the outer limits of profitability for silicon manufacture as things start to get pretty expensive in terms of die yield as you get much above 250mm^2. GPU dies almost always seem to be square as well. Would suggest perhaps an artifical heat die of 16x16mm in size. This would be a useful test tool, not just for GPU blocks, but also for CPU blocks where the present testbed size of 88mm^2 is pretty small in comparison to the ~170-220mm^2 CPU dies that are presently floating about the marketplace. Would be good to have a "small die" and "large die" set of tests, and test all blocks on both to provide a fair idea of the relative strengths in each scenario. Who knows - could find out that a few GPU blocks actually make for some pretty decent small CPU die blocks too.

10-25-2004, 12:05 AM	#10
9mmCensor Cooling Savant Join Date: Apr 2004 Location: Canada Posts: 219	pHaestus maybe contact ATI. They are local (to Canada). Maybe they could sponsor the GPU Cooler testing, you test on their cards, with waterblocks. And because they aren't making waterblocks their is no bais involved. And they get PR.

10-25-2004, 02:59 AM	#11
MaxxxRacer Cooling Savant Join Date: Sep 2004 Location: Los Angeles, Ca Posts: 310	I doubt ati would get anywhere near us. They dont want people frying their cards with water. And this would have them supporting water cooilng their cards. Cool idea though. Would be nice to get pH a free x800xt for his efforts. pH, cathar uses the max OC method to test his cpu blocks. That is why I mentioned it. Between the G4 and G5 he got like 90mhz increase or something... Cant remember exactly what the increase was.. Anyway, it isnt exactly scientific, and to remain credible as you are its best to stick with numbers. As I aksed before, is there anyway to mod ur die simulator to output the right heat and make it approximatly between the die size of a x800 and 6800. right inbetween those. would be better than a 9800 core as it would make the review valid for a longer period of time. __________________ Excuse me, I believe you have my stapler.

10-25-2004, 03:27 AM	#13
MaxxxRacer Cooling Savant Join Date: Sep 2004 Location: Los Angeles, Ca Posts: 310	hrm... well you posted something about getting a higher oc between 2 versions of the storm. you posted on the ausie forums with some pics of cpuz if im not mistaken. correct me if im wrong. This was a while ago btw.. Offtopic again - cathar, thanks for making the storm, you are singlhandedly powering the foward design of waterblocks. (gotta throw some credit over to swifty as they are one of the few desings that didnt rip off cathar) __________________ Excuse me, I believe you have my stapler.

10-25-2004, 02:16 PM	#15
pHaestus Big Player Making Big Money Join Date: Aug 2001 Location: irc.lostgeek.com #procooling.com Posts: 4,782	Yes I am familiar with the ATITOOL for ATI cards. Is there a similar app for NVIDIA cards?

10-31-2004, 02:54 PM	#16
Blackeagle Thermophile Join Date: Oct 2002 Location: U.S.A = Michigan Posts: 1,243	pH, If you go forward with GPU block testing, and I hope you do, could you run the Fusion as one of the early blocks? It's design is one of the more advanced, and it' populer so would give a good start to a graph much like the one in your CPU block articles. I'm sure there are others you wish to do such as the Maze 4 & Swifty designs which would also be good for comparing. I just happen to like the fact that the Fusion has a brass top, so no corrosion risk. And it's the only GPU block I know that offers that, though I may have missed others.

01-22-2005, 04:40 PM	#18
W1zzard Cooling Neophyte Join Date: Sep 2002 Location: Stuttgart, Germany Posts: 25	if you need any customized atitool versions for your testing let me know

03-14-2005, 03:24 AM	#20
MaxxxRacer Cooling Savant Join Date: Sep 2004 Location: Los Angeles, Ca Posts: 310	pH, I am going to contact ATI on your behalf if you dont mind. Also on my behalf i suppose as I will be doing gpu block testing with an x800xt soon enough. __________________ Excuse me, I believe you have my stapler.

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)