- Joined
- Mar 8, 2012
Background:
Long time lurker and first time poster (man I hate when people say that). I have 24x 5970s (48 GPUs) which need to be cooled. I intend to watercool them not for extreme overclocking (they actually will be modestly clocked for stability) but for economics. I have built numerous watercooled systems in the past but nothing on this scale.
The 5970s are dumping 250W of heat ea, combined that's 6 KW (20,000 BTU). This presents two challenges. The first is the GPU farm has been expanded since last summer when a mini split system installed. To handle 6KW I am looking at $3K for upgraded cooling system and then $1500 to $2000 a year in energy costs. So the problem is "simple". I need to move 6 KW of heat from inside to outside using liquid and if I can do it under $5K it will pay for itself in a year. Later I also plan to tap some of those BTUs to preheat hot water and possibly heat the house in winter.
The details:
The 24x HD 5970s GPUs are currently air cooled 3 graphics cards (6 GPUs) per system. I will be moving them to 4x5970 configuration in a 4U chassis. Yes 48 GPUs in 24U of rack space. To anyone wondering the servers are controlled by a non-GPU "watchdog" server which configures, monitors, and controls them (including power cycle and power off). The watchdog will be expanded to provide monitoring on the entire cooling loop (flow, pump current, ambient air temp, reservoir level, 2 water temp sensors, and 2 outside air temp sensors.
Here is a test system. Excuse the ugly routing. This was more to test that 4x5970s can run under load 24/7 in a 4U rackmount chassis. It has been running like this under 100% load for a week now.
The test rig has an internal MCP655-B and 5.25" bay reservoir. Initially I had considered having a "closed" internal loop (4 GPU -> pump -> flat plate heat exchangers -> 4 GPUs) and link the cold side of flat plat heat exchanger to an outer loop. This would provide fluid isolation (sealed inner loop for each chassis & shared outer loop linking all the chassis). While I still think that is optimal it is simply too expensive.
The revised plan is to have a mainline and then a manifold to run branch lines to each rig. Each rig connects to the hot-side and cold-side manifolds via flexible tubing and quick disconnect fittings. The mainline will connect to a large water to air heat exchanger and 16" 1000cfm fan mounted outside.
ASCII diagram might help:
Here is rig connected to a test radiator. Yeah 4x120mm is undersized for 1000W thermal load but it being winter helped. Once again it was more a proof of concept to test it before building the system out.
Hey noob didn't you say something about a pump?
Right I just thought some info on the setup might help.
So I could use the help of experts on a couple points.
Pump:
So I (maybe incorrectly see below) am looking for a pump capable of 12 gpm (16 preferably) at a head of 8 feet. Totally wrong? You tell me. Heat dump is less of an issue than in most setups as it is dwarfed by the 6KW being transfered from the GPUs. The problem is that excludes all "normal" water cooling pumps. They are simply not even in the ballpark. Even dual high end pump solutions fall very short. I first looked a radiant heat circulation pumps but most of them are cast iron. Sump pumps might work but I am not sure if they are designed for 24/7 operation. Most aquarium pumps lack the flow I am looking for.
I then looked at Iwaki RD series but the RD-30 only has max GPM of ~6 which is insufficient. The RD-40x might be useful (18 gpm @ 4ft, 16 gpm @ 8 ft) but I can't find it anywhere. Iwaki does make another series, the MD (and WMD). They seem to be a good fit. They have 1" NPT fittings, high head, and high flow.
http://www.iwakiamerica.com/products/wmd.htm
looks like MD is made in Japan and WMD is made in the US.
I was thinking of a WMD-30RXLT or maybe WMD-40RXLT but I thought I would check here.
Flow necessary:
Since there will be 6 servers (w/ space for 8) and they will be connected to manifold in parallel to acheive 2gpm across each server would require a pump capable of 16gpm (at whatever head the system creates). Right?
Parallel vs Serial:
My test rig has the 4 GPUs in parallel. Likely I shouldn't right? Flow is reduced across each GPU. i.e. if flow into SLI block is 4 GPM then when it branches parallel 4 ways it is 1 GPM across each GPU. I understand normally this won't matter but remember each server will also
be in parallel.
Head:
I have no idea how much head would be necessary for a system like this. I also have no idea how to calculate it for one server and if I knew what it is for 1 server is it the same for 6/8 servers (because they are in parallel)? Any insight, thoughts, guesstimates, or tips would be helpful.
Thoughts?
I am open to any other pump recommendations. Also some confirmation/correction on the likely flow and head I am looking at would be nice. Yeah I know it was long but I wanted to avoid a "get a Liang" or "aquarium pumps" suck type response.
Long time lurker and first time poster (man I hate when people say that). I have 24x 5970s (48 GPUs) which need to be cooled. I intend to watercool them not for extreme overclocking (they actually will be modestly clocked for stability) but for economics. I have built numerous watercooled systems in the past but nothing on this scale.
The 5970s are dumping 250W of heat ea, combined that's 6 KW (20,000 BTU). This presents two challenges. The first is the GPU farm has been expanded since last summer when a mini split system installed. To handle 6KW I am looking at $3K for upgraded cooling system and then $1500 to $2000 a year in energy costs. So the problem is "simple". I need to move 6 KW of heat from inside to outside using liquid and if I can do it under $5K it will pay for itself in a year. Later I also plan to tap some of those BTUs to preheat hot water and possibly heat the house in winter.
The details:
The 24x HD 5970s GPUs are currently air cooled 3 graphics cards (6 GPUs) per system. I will be moving them to 4x5970 configuration in a 4U chassis. Yes 48 GPUs in 24U of rack space. To anyone wondering the servers are controlled by a non-GPU "watchdog" server which configures, monitors, and controls them (including power cycle and power off). The watchdog will be expanded to provide monitoring on the entire cooling loop (flow, pump current, ambient air temp, reservoir level, 2 water temp sensors, and 2 outside air temp sensors.
Here is a test system. Excuse the ugly routing. This was more to test that 4x5970s can run under load 24/7 in a 4U rackmount chassis. It has been running like this under 100% load for a week now.
The test rig has an internal MCP655-B and 5.25" bay reservoir. Initially I had considered having a "closed" internal loop (4 GPU -> pump -> flat plate heat exchangers -> 4 GPUs) and link the cold side of flat plat heat exchanger to an outer loop. This would provide fluid isolation (sealed inner loop for each chassis & shared outer loop linking all the chassis). While I still think that is optimal it is simply too expensive.
The revised plan is to have a mainline and then a manifold to run branch lines to each rig. Each rig connects to the hot-side and cold-side manifolds via flexible tubing and quick disconnect fittings. The mainline will connect to a large water to air heat exchanger and 16" 1000cfm fan mounted outside.
ASCII diagram might help:
Code:
3/8" PEX branch lines
from heat-exchanger
|--------{server #1}--------|
1" PEX cold line | | 1" PEX hot line
-------------------------|--------{server #2}--------|----------------- to pump/reservoir/heat-exchanger --->
| |
|--------{server #3}--------|
cold side manifold hot side manifold
-----(flow direction)--------->
Only 3 servers show for simplicity. Total of 6 will be used.
Here is rig connected to a test radiator. Yeah 4x120mm is undersized for 1000W thermal load but it being winter helped. Once again it was more a proof of concept to test it before building the system out.
Hey noob didn't you say something about a pump?
Right I just thought some info on the setup might help.
So I could use the help of experts on a couple points.
Pump:
So I (maybe incorrectly see below) am looking for a pump capable of 12 gpm (16 preferably) at a head of 8 feet. Totally wrong? You tell me. Heat dump is less of an issue than in most setups as it is dwarfed by the 6KW being transfered from the GPUs. The problem is that excludes all "normal" water cooling pumps. They are simply not even in the ballpark. Even dual high end pump solutions fall very short. I first looked a radiant heat circulation pumps but most of them are cast iron. Sump pumps might work but I am not sure if they are designed for 24/7 operation. Most aquarium pumps lack the flow I am looking for.
I then looked at Iwaki RD series but the RD-30 only has max GPM of ~6 which is insufficient. The RD-40x might be useful (18 gpm @ 4ft, 16 gpm @ 8 ft) but I can't find it anywhere. Iwaki does make another series, the MD (and WMD). They seem to be a good fit. They have 1" NPT fittings, high head, and high flow.
http://www.iwakiamerica.com/products/wmd.htm
looks like MD is made in Japan and WMD is made in the US.
I was thinking of a WMD-30RXLT or maybe WMD-40RXLT but I thought I would check here.
Flow necessary:
Since there will be 6 servers (w/ space for 8) and they will be connected to manifold in parallel to acheive 2gpm across each server would require a pump capable of 16gpm (at whatever head the system creates). Right?
Parallel vs Serial:
My test rig has the 4 GPUs in parallel. Likely I shouldn't right? Flow is reduced across each GPU. i.e. if flow into SLI block is 4 GPM then when it branches parallel 4 ways it is 1 GPM across each GPU. I understand normally this won't matter but remember each server will also
be in parallel.
Head:
I have no idea how much head would be necessary for a system like this. I also have no idea how to calculate it for one server and if I knew what it is for 1 server is it the same for 6/8 servers (because they are in parallel)? Any insight, thoughts, guesstimates, or tips would be helpful.
Thoughts?
I am open to any other pump recommendations. Also some confirmation/correction on the likely flow and head I am looking at would be nice. Yeah I know it was long but I wanted to avoid a "get a Liang" or "aquarium pumps" suck type response.
Last edited: