TriFire water cooling (was: Squeezing 2-slot cards into a single slot)

[DeathAndTaxes]

i have 6 6970's watercooled side by side and works great.

How do you get around the 8 GPU limit in AMD drivers?

He doesn't - 6970 is a single-GPU card.

DOH.  I could have sworn he said 5...970.

MadHacker grab some photos.  6 watercooled GPU is worth sharing.

[1713306179]

1713306179

[1713306179]

1713306179

[MadHacker]

I use these Feser Crossfire / SLI Multi Spacing Fittings
http://www.frozencpu.com/products/8566/ex-tub-448/Feser_Crossfire_SLI_Multi_Spacing_Fittings.html?tl=g30c101s873#blank
as the poster above sugested.

i have 6 6970's watercooled side by side and works great.

How do you get around the 8 GPU limit in AMD drivers?

6970's are only 1 GPU, your thinking of 6990's or 5970's
so i'm only using 6 GPUs total.
having a real problem getting 7 GPUs working in a single machine

[cicada]

I use these Feser Crossfire / SLI Multi Spacing Fittings
http://www.frozencpu.com/products/8566/ex-tub-448/Feser_Crossfire_SLI_Multi_Spacing_Fittings.html?tl=g30c101s873#blank
as the poster above sugested.

Mad, are you doing serial (in one card, out to the next) or parallel (in all, out all)?

[MadHacker]

I use these Feser Crossfire / SLI Multi Spacing Fittings
http://www.frozencpu.com/products/8566/ex-tub-448/Feser_Crossfire_SLI_Multi_Spacing_Fittings.html?tl=g30c101s873#blank
as the poster above sugested.

Mad, are you doing serial (in one card, out to the next) or parallel (in all, out all)?

parallel.
SideWinder had a deal on the Crosfire/SLI fittings so i bought 12 of them for $4 each.
they work great.

[DeathAndTaxes]

I use these Feser Crossfire / SLI Multi Spacing Fittings
http://www.frozencpu.com/products/8566/ex-tub-448/Feser_Crossfire_SLI_Multi_Spacing_Fittings.html?tl=g30c101s873#blank
as the poster above sugested.

Mad, are you doing serial (in one card, out to the next) or parallel (in all, out all)?

You don't want to do serial if each waterblock is the same.  The water temp rises across each GPU load.  The last card gets the hottest coolant and yet you will still expect the same performance from it.  Parallel is much better.  The inlet temp for each waterblock remains the same.

IF you have dissimilar blocks (like GPU block and CPU block) you need to do serial (or separate loops) because otherwise water will "cheat".  If one block has higher resistance the water will flow around it through the other block which is not good.  If you decide to water cool your CPU also make sure it is before the GPU.  The GPU have higher thermal load and thus will warm the coolant more.  CPU are also less accommodating of high core temps than GPU are.

[cicada]

Am I correct that I only need 6 fittings for 4 blocks then?



IN - 
     ----------- (gpu1)
       |      |
     ----------- (gpu2)
       |      |
     ----------- (gpu3)
       |      |
     ----------- (gpu4)
                 - OUT     
             

Where each | is an SLI fitting?

[MadHacker]

Am I correct that I only need 6 fittings for 4 blocks then?

IN - 
     ----------- (gpu1)
       |      |
     ----------- (gpu2)
       |      |
     ----------- (gpu3)
       |      |
     ----------- (gpu4)
                 - OUT     
             

Where each | is an SLI fitting?

that is correct.

[DeathAndTaxes]

Am I correct that I only need 6 fittings for 4 blocks then?



IN - 
     ----------- (gpu1)
       |      |
     ----------- (gpu2)
       |      |
     ----------- (gpu3)
       |      |
     ----------- (gpu4)
                 - OUT     
             

Where each | is an SLI fitting?

I found this is easier to route due to power supply and case location.  It is also easier to drain and disconnect as both of the inlets to the GPU cluster are on the same waterblock.

Code:

      IN    OUT
       |      | 
       -->------- (gpu1)
       |      |
       -->------- (gpu2)
       |      |
       -->------- (gpu3)
       |      |
       -->------- (gpu4)
 

I added an arrow > on each water block to indicate that the water flows across each water block.

You just don't want to do this

Code:

      IN   
       |      
       -->------- (gpu1)
       |      |
       -->------- (gpu2)
       |      |
       -->------- (gpu3)
       |      |
       -->------- (gpu4)
       |
    OUT 

That will cook the GPU in seconds after booting. 

[BTCurious]

I found this is easier to route due to power supply and case location.  It is also easier to drain and disconnect as both of the inlets to the GPU cluster are on the same waterblock.

Code:

      IN    OUT
       |      | 
       -->------- (gpu1)
       |      |
       -->------- (gpu2)
       |      |
       -->------- (gpu3)
       |      |
       -->------- (gpu4)
 

I added an arrow > on each water block to indicate that the water flows across each water block.

This is not entirely equal, your gpu 4 will get less flow. Although if your interconnects don't pose a lot of friction, it shouldn't be much of a problem.

[DeathAndTaxes]

This is not entirely equal, your gpu 4 will get less flow. Although if your interconnects don't pose a lot of friction, it shouldn't be much of a problem.

True however flow usually isn't the limit of cooling capacity.   It takes a surprising little amount of flow to remove even 1 KW thermal load.  As long as your flow is high enough that it doesn't result in laminar flow you will be fine.  To anyone who doesn't know laminar flow is when the water becomes "smooth" and "efficient".  We don't want that because then only heat is transfered to the edge of the water stream reducing cooling capacity.  Modern waterblocks tend to create enough "chaos" that laminar flow won't happen unless flow rate is very very low. 

Most pumps have more than enough head to keep flow rate > 1 gpm which is more than enough to avoid laminar flow.

[MadHacker]

I found this is easier to route due to power supply and case location.  It is also easier to drain and disconnect as both of the inlets to the GPU cluster are on the same waterblock.

Code:

      IN    OUT
       |      | 
       -->------- (gpu1)
       |      |
       -->------- (gpu2)
       |      |
       -->------- (gpu3)
       |      |
       -->------- (gpu4)
 

I added an arrow > on each water block to indicate that the water flows across each water block.

actual i have the same set up as you have listed above.
checking temps they are 45-53 degC
ambiant temp is about 28Deg.
6970 OC to 965, getting 432MH/s

[RandyFolds]

Before you fill your reservoirs, be sure to read this: https://bitcointalk.org/index.php?topic=40325.0;topicseen  

MadHacker, pics pics pics!

There are a lot of aquarium plumbing solutions that may help for these crazy custom configurations. They are also probably a fair bit cheaper for Guest fittings and valves and the like. You can also find some crazy gang-valves that will allow you to tailor the flow to each exchanger. www.marinedepot.com is a pretty good source to at least find pics of it...local is always faster and cheaper.

[DeathAndTaxes]

Before you fill your reservoirs, be sure to read this: https://bitcointalk.org/index.php?topic=40325.0;topicseen  

OH MAN.  Thanks for the laughs.  I love how you got "Bonkers" pissed off because he thought you were serious about the gravy.

Honestly people love to over think it.  There are no low cost fluids that have higher thermal conductivity than distilled water.  Distilled water is cheap, has low conductivity, high thermal conductivity, is easy to pump, and is non-reactive.  Kinda hard to beat that ... well except maybe with Gravy.

[MadHacker]

Before you fill your reservoirs, be sure to read this: https://bitcointalk.org/index.php?topic=40325.0;topicseen  

OH MAN.  Thanks for the laughs.  I love how you got "Bonkers" pissed off because he thought you were serious about the gravy.

Honestly people love to over think it.  There are no low cost fluids that have higher thermal conductivity than distilled water.  Distilled water is cheap, has low conductivity, high thermal conductivity, is easy to pump, and is non-reactive.  Kinda hard to beat that ... well except maybe with Gravy.

unfortunatlty distiled water startes out not very conductive... but over time it picls up ions from the metal in the blocks...
then 1 leak and all can be toast...

I have RMA'd quite a few motherboards in the past because they couldn't swim (of course i didn't tell Asus that)

as for Picts... I'll try to get some taken over the weekend.

[RandyFolds]

I am pretty sure Bonkers was a troll from SomethingAwful, but he/she sure succeeded in getting some people worked up in very short order.

[cicada]

Just ordered all the parts to start this up, thanks for all the tips guys

[ssateneth]

Why not like this?


      IN   
       |       
       -->------- (gpu1)
              |
       --<------- (gpu2)
       |     
       -->------- (gpu3)
              |
       --<------- (gpu4)
       |     
      OUT

[cicada]

Why not like this?

I did a *lot* of reading and research last night, the consensus seems to be that it doesn't matter too much, but parallel flow through the gpu array improves overall flow rates through any other blocks in your system,  usually the CPU.

[BTCurious]

Ḃecause then gpu 4 will only get warm fluid, and run much hotter than the rest.

[DeathAndTaxes]

Why not like this?


      IN   
       |       
       -->------- (gpu1)
              |
       --<------- (gpu2)
       |     
       -->------- (gpu3)
              |
       --<------- (gpu4)
       |     
      OUT

If each GPU caused a 1 deg rise in water temp then the inlet temp on GPU4 would be 3 deg higher.  It likely doesn't matter because water is such an effective heat transfer fluid that all the water in entire loop is within 1 to 4 degrees of water anywhere else in the loop.

Still parallel is considered a better approach.  Also with your layout you have 6 right turns.  Ever 90 turn is roughly equal to another 2 feet of tubing in terms of the resistance it creates.  Once again pumps today are so powerful and min effective flow rates so low it is likely academic but generally a parallel approach w/ minimum number of 90deg turns is the best.