Stren's Titan Water Block Roundup

[tiborrr]

http://www.hardwaremax.net/wasserkue...owall=&start=8

[stren]

http://www.hardwaremax.net/wasserkue...owall=&start=8

Hey Niko - yeah I hear you, I was surprised too. Here are some differences between the two setups:

1. Load - hardwaremaxx used a small 15% overclock and ran heaven. Neither the benchmark or overclock is as power intensive as my setup where I overclock to 1123MHz (34% increase) and run furmark in extreme burn mode which produces a consistent 124% power load, far above what a card with an unmodified bios could do (106% max). I used a fixed voltage and frequency and did not allow the card to deviate from these settings like it normally would.

2. Throttling Fix - I see no mention of logging gpu frequency and voltage or an upgraded bios to ensure that there was no throttling going on. Throttling is a well known and big issue on many (though not all) Titans.

3. Fixed flow vs variable - the hardwaremaxx data is run at a single pump setting - I calculated it out and my nearest data point is the "129" pump setting. Here I also see EK and Aquacomputer much closer together than HardwareMaxx did. Possibly because of number 4.

4. TIM - It's not clear to me whether HardwareMaxx used the supplied TIM or MX2. MX2 is specified under the hardware system setup though that could be in regard to the CPU, it's never specified that this was used for the test. It's perfectly possible that XSPC's supplied TIM was used and is pretty crappy or has a long curing time.

5. Multiple mounts vs One - My data is only one mount so far, hardwaremaxx completed three mounts and took the best result. Maybe EK had a bad first mount although the TIM spread looked decent?

6. Temperature sensor location - from the picture of the setup the temperature sensor is not located close to the EK titan. I mount my coolant temperature sensor immediately next to the inlet of the GPU to ensure accurate coolant readings.

BTW will the SE block have any different performance? If anyone has any further ideas or information then please let me know, I'm all about making the best test possible. I'm not trying to discredit hardwaremaxx's review but merely to try and understand why the data might be different. If anything I'm surprised their data spread was so large - in their testing the AquaComputer block cooled twice as well as the XSPC, while in mine it performed ~15% worse. Given what we've seen in CPU block testing a performance change of 15% seems a lot more reasonable than 100%. I can change the way I test if there is a better way, but I can't change the numbers that I get back from the testing Well I could, but I don't :p

I also got done with taking the "VRM"/"VRAM" temperature measurements on two blocks. These are measured with an uncalibrated IR temperature sensor vs ambient temps. Measuring vs ambient is far from ideal. Also taking measurements at one point in time is also far from ideal as is measuring the back of the board underneath the components. Essentially take these numbers with a grain of salt, and look for big differences only:



Here we can see VRAM temps are essentially the same for the quality of the test. However we see a decent improvement on the VRMs with the XSPC block. I need to take the EK measurements also as I didn't have the IR thermometer for that point in time. Given that this seems like a reasonable way to measure things even if the method is far from the best, I'll be investing in some more temperature probes to stick to the back of the board for more accurate measurements.

I also took some photos of the xspc block on the card after testing. My normal photo editing rig is down as the PSU died, so the border etc isn't consistent





I then attached the HydroCopper block. There were 5 different types of screws of which the instructions mentioned 2 different types... The screws are also horribly small and easily strippable. By far the worst block install so far. Leak testing going on now. Results most likely on saturday unless I get busy on the weekend in which case monday

[tiborrr]

Today I quickly ran our GTX Titan card and the highest delta I saw was about 8.5-9?C (D5 on setting 1) or about 6.5-7?C (D5 on setting 5). The loops consists of a single 360 rad and a D5 X-RES system. Even if the block contact was less than perfect one would suspect the higher mass flow cannot solve this. Therefore, a delta of 2K should be expected but not the delta of 10K.

Because the BIOS is still the stock one I could overclock it by +150 / +450 / 106% (GPU/RAM/Power limit respectively). If you send it to me (via email) I can try to replicate the results on Monday.

What I find odd is:
1. EK-FC Titan scaling so well with high flow rates. Usually, low-restriction water blocks like ours do not gain much with higher flow rates. On the contrary, blocks with exceptionally narrow micro-channel structure (e.g. Koolance, Aquacomputer) react positively performance-wise to higher flow rates. It is simple logic really - the more surface area, the bigger the effect. I believe there is a problem with (consistent) throttling of your GTX Titan or maybe with the thread lenght on your fittings? The fitting thread lenght should NOT be longer than 5mm. See this picture: http://i.imgur.com/PjBi23mh.jpg
2. XSPC block is way off the charts. It has the smallest cooling engine of them all therefore it should perform even worse with larger heat loads (given the contact is okay with all of the tested water blocks).
3. Every block you test it performs better. 10?C on a VRM is a too big of a margin given the fact both XSPC & Aquacomputer blocks have direct contact with VRMs. Where and by using what methodology (IR, wire sensor, foil sensor) is that measurement taken?- NEVERMIND I RE-READ YOUR POST. Do you think there is a chance your temperature measuring equipment might be the case as well?

Talk to you soon,
Niko

[stren]

Today I quickly ran our GTX Titan card and the highest delta I saw was about 8.5-9?C (D5 on setting 1) or about 6.5-7?C (D5 on setting 5). The loops consists of a single 360 rad and a D5 X-RES system. Even if the block contact was less than perfect one would suspect the higher mass flow cannot solve this. Therefore, a delta of 2K should be expected but not the delta of 10K.

Yeah my pump range is far wider than a D5 vario can handle though. What flow rates were you getting? If you look at 0.6GPM to 1.3GPM range which is probably similar to your D5 then my block also varies ~3C, the difference is at the very low and very high flow rates. Bear in mind I use a MCP35x2 pump. My lowest flow rate is with PWM at 23% (RPM range is pretty linear from ~19%-51%) and only powering one pump, max flow is with both pumps running at max.

Because the BIOS is still the stock one I could overclock it by +150 / +450 / 106% (GPU/RAM/Power limit respectively). If you send it to me (via email) I can try to replicate the results on Monday.

Yeah I don't trust the stock bios at all. The rig I flashed on is down right now, but here's the link:

http://1pcent.com/files/nvidia/121nv150.zip

I didn't overclock the memory at all, but I ran 1.212/1123 with the option to lock voltages and frequencies. I forget what it's called in precision-x to do that, but I couldn't find the same thing in afterburner so I gave up on afterburner.

What I find odd is:
1. EK-FC Titan scaling so well with high flow rates. Usually, low-restriction water blocks like ours do not gain much with higher flow rates. On the contrary, blocks with exceptionally narrow micro-channel structure (e.g. Koolance, Aquacomputer) react positively performance-wise to higher flow rates. It is simple logic really - the more surface area, the bigger the effect. I believe there is a problem with (consistent) throttling of your GTX Titan or maybe with the thread lenght on your fittings? The fitting thread lenght should NOT be longer than 5mm. See this picture: http://i.imgur.com/PjBi23mh.jpg

Yup this was very strange to me too, we've seen some blocks in the cpu roundup that scaled a bit better than others but this much scaling was very unprecedented. The fittings I use are the bitspower Q fitting on the inlet side, and the koolance g1/4 to g3/8 adapter on the outlet port (on the thicker side of the bridge). I'll double check the length but I think they are pretty short. Bear in mind the flow rates were still the best out of any block so this is unlikely to be the cause.

2. XSPC block is way off the charts. It has the smallest cooling engine of them all therefore it should perform even worse with larger heat loads (given the contact is okay with all of the tested water blocks).

Yeah I haven't opened up the block to measure the fin area and size so I can't confirm this, but like I said it was a surprise.

3. Every block you test it performs better.

Yeah the retests will show if there's some kind of "burn in" process for the card that might be causing performance to shift with time.

10?C on a VRM is a too big of a margin given the fact both XSPC & Aquacomputer blocks have direct contact with VRMs. Where and by using what methodology (IR, wire sensor, foil sensor) is that measurement taken? Do you think there is a chance your temperature measuring equipment might be the case as well?

Talk to you soon,
Niko

Like I said this is measured on the back of the board directly below where the VRMs are using an IR "laser" thermometer. This is the first time that I've used this tool, and will be switching over to real probes soon for hopefully better measurements. There's a high degree of error, however I doubt it's 10 degrees big. Thermal pads used are different thicknesses and materials and so some difference would be expected. I'm surprised the VRM area gets this hot under any full cover block to be honest. If the board is this hot, presumably the junction temp of the VRM itself is hotter unless the board is actually getting heated by a nearby device that isn't cooled by the block.

[tiborrr]

Love your replies, thanks for a quick response!

[stren]

Love your replies, thanks for a quick response!

NP - I love to see correlation too, but at the same time I think most manufacturer's expect their block to be top because they've designed it for their setup and not my setup. At the same time I like my testing to be representative of an end user's setup which means trying to correlate as best I can, so this kind of discussion is healthy and helps towards that goal.

Looking through the data, the only other thing is that the EK block's ambients during testing were in general lower than the others and this may have some small second order effect due to possibly poor linearity of temperature sensors in the card. So when I retest I'll try and make sure the garage gets nice and toasty.

[paulbagz]

Hey stren, will you be able to get your hands on the Swiftech Komodo?

-PB

[tiborrr]

Hey stren, will you be able to get your hands on the Swiftech Komodo?

-PB

Swiftech Komodo = EVGA Hydrocopper.

[stren]

Swiftech Komodo = EVGA Hydrocopper.

This is true. First data point on the block (high flow only) is pretty standard, all blocks so far (excluding Koolance) have been between 7.3 and 7.9C delta at max flow. We'll have to see how it responds to lower flows though. It's also seems to be equaling EK for the least restrictive block. As for PCB temps under the VRMs it's the worst yet: ~60C above ambient!

[paulbagz]

Swiftech Komodo = EVGA Hydrocopper.

Indeed, hence the desire to see results to see if the price tag warrants it or not.

-PB

[cr1t1cal]

Hydrocopper GTX 680 160$ waterblock structure.



Hope to see the Titan model.

[stren]

Indeed, hence the desire to see results to see if the price tag warrants it or not.

-PB

Hopefully hydrocopper testing will be done tomorrow, but i may not get done with analyzing it until monday. Swiftech are sending a Komodo too though the results should of course be identical. I may run that one last because of that.

[dengyong]

I have no numbers available. So far we only did a quick test to check if the heatpipe transports heat and it does

Today the card have met his new friends:

That's some beautiful work !.. Will your nickle/plexi support distilled water with a small trace of copper sulfate as a biocide ?

[lowfat]

Hmm. With previous generation cards, Swiftech never sold Komodos (for Nvidia) under their own brand. It was just the Hydrocopper. That is IMO by far the best looking Swiftech block in years. Even the backplate looks good.

[Vetalar]

@lowfat
what 'bout ATI fullcovers - they were branded as comodo and AFAIR sold only separately

[lowfat]

@lowfat
what 'bout ATI fullcovers - they were branded as comodo and AFAIR sold only separately

Sorry I meant to say for Nvidia cards.

[stren]

Hmm. With previous generation cards, Swiftech never sold Komodos (for Nvidia) under their own brand. It was just the Hydrocopper. That is IMO by far the best looking Swiftech block in years. Even the backplate looks good.

Yeah this has been the case in the past, but with the H220 success there is more demand for swiftech branded blocks and of course EVGA still want their own branded blocks too. The only differences should be the logos and the LED color. However the Swiftech backplate is different to the EVGA backplate, each one is manufactured by their respective company. Gabe was saying that their backplate is hella beefy 6mm aluminum. I'll see how it compares when it comes in.

[Martinm210]

Today I quickly ran our GTX Titan card and the highest delta I saw was about 8.5-9?C (D5 on setting 1) or about 6.5-7?C (D5 on setting 5). The loops consists of a single 360 rad and a D5 X-RES system. Even if the block contact was less than perfect one would suspect the higher mass flow cannot solve this. Therefore, a delta of 2K should be expected but not the delta of 10K.

Because the BIOS is still the stock one I could overclock it by +150 / +450 / 106% (GPU/RAM/Power limit respectively). If you send it to me (via email) I can try to replicate the results on Monday.

What I find odd is:
1. EK-FC Titan scaling so well with high flow rates. Usually, low-restriction water blocks like ours do not gain much with higher flow rates. On the contrary, blocks with exceptionally narrow micro-channel structure (e.g. Koolance, Aquacomputer) react positively performance-wise to higher flow rates. It is simple logic really - the more surface area, the bigger the effect. I believe there is a problem with (consistent) throttling of your GTX Titan or maybe with the thread lenght on your fittings? The fitting thread lenght should NOT be longer than 5mm. See this picture: http://i.imgur.com/PjBi23mh.jpg
2. XSPC block is way off the charts. It has the smallest cooling engine of them all therefore it should perform even worse with larger heat loads (given the contact is okay with all of the tested water blocks).
3. Every block you test it performs better. 10?C on a VRM is a too big of a margin given the fact both XSPC & Aquacomputer blocks have direct contact with VRMs. Where and by using what methodology (IR, wire sensor, foil sensor) is that measurement taken?- NEVERMIND I RE-READ YOUR POST. Do you think there is a chance your temperature measuring equipment might be the case as well?

Talk to you soon,
Niko

I disagree. I have typically seen blocks that are higher in restriction has less sensitivity to flow rate changes and generally perform better at low flow rates. The more micro pin/fin whatever the smaller the net total cross section which will net higher velocity. Of coarse it is a balancing act with loss in overall net flow rate loss due to the restriction loss.

Take a look at the DT5noz in Strens awesome CPU block work. Vey much an extremely low restriction block with it's parallel nozzle design and also the most gained with increased pumping power.



That has been fairly typical of restriction vs flow sensitivity. Not always but generally what I have seen.

As always I would encourage manufactures to share their own. The more data and the more test benches/conditions the better.

Thanks for all your hard work Stren!

IMHO it seems GPU performance is so darn close here that perhaps VRM cooling is the more important performance differentiator. I know nothing about the card, but I have heard some others GPUs being more sensitive to VRM cooling

[stren]

IMHO it seems GPU performance is so darn close here that perhaps VRM cooling is the more important performance differentiator. I know nothing about the card, but I have heard some others GPUs being more sensitive to VRM cooling

Very good points on the 5Noz Martin, although the MIPS block was also massively parallel and didn't see the same effect. Whenever I try and read patterns in my data I always just end up confused lol.

Yeah at some point you feel like whether the core is 9C or 10C above coolant doesn't make a big deal of difference, but 40C vs 60C above ambient on VRMs might.

I actually got done with the hydrocopper on saturday morning, but my workstation pump died yesterday so I had to strip down my TX10 and built a quick bench rig to replace it (I wish I had another of spotswood's tech benches, but the CL motherboard tray doubles decently). Here's some data on the evga/swiftech block:



Decently performing block on the core, very low restriction:



Pretty horrible VRM temps, VRAM temps seems all about the same:

[stren]

So I ran the 2nd Koolance block without the backplate using the "GTX780 Screw Kit". I used the suggested thermal pads and the TIM spread looked a lot better this time. I did end up with what I would consider unacceptable warping in the card however, so I'd like to re run to see if I can maintain the performance without putting stress on the card. This run at least shows the best possible performance and then we can back off a bit. Luckily the warp did not kill the card (yet). Next up is the EK re run. EK's TIM spread while better than the original Koolance still looked a little thick. I'm going to try and tighten it as best I can without bending the card. Results below:



Again VRM/VRAM temps have a much higher degree of error:

[paulbagz]

Go XSPC!

-PB

[Martinm210]

Nice work! Looks like VRM is the key differentiator. XSPC looking like the better performer.

Don't like hearing about card warpage on the koolance block. I wouldn't push it, that card is worth way to much to kill for some additional numbers. I doubt anyone in regular use would purposely warp their brand new titan either just to get an extra half a degree.

[stren]

Nice work! Looks like VRM is the key differentiator. XSPC looking like the better performer.

Don't like hearing about card warpage on the koolance block. I wouldn't push it, that card is worth way to much to kill for some additional numbers. I doubt anyone in regular use would purposely warp their brand new titan either just to get an extra half a degree.

I agree - the warp was disconcerting but I wanted to see if with good contact that the block itself was good, which it is. I'm certainly not trying that again lol. Next up will be seeing how this 2nd block performs without warping. My gut feel is that the EK and Koolance blocks don't have quite the right height of the gpu core, EK is off a tiny bit so gets decent but not great TIM spread, and that the Koolance is off by a lot. Naennon solved this by removing the thermal pads, but the real solution maybe to alter the block design a bit.

But yes the core temps are so similar that the VRM temps seem to be the biggest factor. I also think that the biggest factor in my measurement of the VRM temps is the heatsinking on what I think are the "current sense resistors" next to the VRMs themselves. This is because I measure the back of the PCB under the VRMs. Take the Hydrocopper block for example even when I added thermal pads to those resistors they still didnt even make contact with the block which is I think why it's results were so bad. Even though I'm not so much measuring the VRM temp as the board temp, a hotter board will mean the VRM itself runs hotter too. Eitherway if I'm using a full cover waterblock my expectation would be that the board doesn't hit 96C!

[Razor Time]

Thank you Stren for taking the time to test and report on these blocks, always great to see info like this when buying gear.

I can't seem to get my head around these vrms temps and the vast difference in results. Surely if contact is being made with the vrms then we shouldn't be seeing high temps like we are with some of these blocks?

[stren]

Thank you Stren for taking the time to test and report on these blocks, always great to see info like this when buying gear.

I can't seem to get my head around these vrms temps and the vast difference in results. Surely if contact is being made with the vrms then we shouldn't be seeing high temps like we are with some of these blocks?

I honestly think it's more to do with contact on the resistors next to the VRMs and not all blocks do make good contact with those. I also think EK uses unusually thin thermal pads which explains it's results below. For example the swiftech didn't make contact with the resistors even though I added extra thermal pads on those resistors. Which leads me to the next question? My new probes came in so I can measure and log tons of points on the back of the board now. So far I have always added thermal pads to the resistors regardless of the installation instructions. Should I continue this method or follow the manufacturer's instructions. I.E. Do what I think is best, or punish those manufacturer's who aren't as smart as me (jk jk)?

Here's EK's VRM and VRAM temps. Very surprising.