Is LN2 on GPU really worth the trouble?

  • News, Editorials
  • 23

Is LN2 on GPU really worth the trouble?

Author: Pieter-Jan Plaisier

Extreme overclocking is not only about using as much liquid nitrogen as possible to get the best results, it’s also about knowing if it’s interesting to use another batch of 50L LN2 or if the hardware has been maxed out completely. In addition, for the not-so-extreme overclockers who don’t have the resources to acquire liquid nitrogen every weekend there’s the problem of how much switching to LN2 will actually improve the score. The HWBOT database seems to have an answer to these questions …

First of all, being an extreme overclocker (and extreme tech enthusiast) myself, I have to say the following: “YES, it is always interesting to test LN2”. Whether it’s a high-end Radon 5870 or a low-end Geforce GT220 … liquid nitrogen can give you a better view on the scalability of you hardware.

Now, to continue finding an answer to the questions posed in the first paragraph. The basic question we posed here is: “To what extend is changing the temperature of the hardware configuration yielding higher performance”. So, for each video card, we have two variables; the independent variable ‘temperature’ and the dependent variable ‘performance’ or, to make it a bit easier to understand, we manipulate the temperature and measure the performance. The more the performance increases with given decrease in temperature, the more interesting it is to use liquid nitrogen to cool down the hardware. All agree?

Now, as I prefer to follow a scientific methodology as much as possible, I would normally assemble a couple of test systems and do the testing myself. However, the main issue we face here is the lack of time and finances for this kind of semi-academic purposes, so instead I’m using the 400,000 results counting HWBOT database. Because time is actually a very-much restraining factor, I have also limited the measurement tool (performance) to 3DMark06, single card as well as the amount of temperature variables:

# +°C CPU/GPU = positive temperatures for both CPU and GPU, meaning AIR or WATER cooled configurations
# -°C CPU / +°C GPU = subzero temperature for CPU, but still an air/water cooled GPU
# -°C CPU/GPU = subzero configuration, so both CPU and GPU below 0°C

As for the hardware, I have limited the amount of cards to the top 8 cards of ATI and Nvidia, excluding those samples that don’t have enough data to fit all the temperature categories (e.g.: Geforce GTX 470). Another important side note is that I have no control over the used test platform, which means that the individual results may vary because different platforms have been used. Do note, however, that this doesn’t make the charts invalid as we are not focusing on the performance of the hardware in comparison to each other, but the performance of the hardware compared based on temperature.


In this first chart, which is solely used to display what we’re measuring here, we can see that in general it’s always interesting to decrease the temperature of a given platform as the performance always increases when we change cooling methods. A trained eye, however, will notice instantly that the difference between cooling methods differs from hardware to hardware. A simple example is the difference between the Radeon 5970 scaling and Geforce GTX 275.

To find out what video card is most interesting to blow your LN2 on, we have to alter our data a bit to see the percentual increase over air/water cooled configurations.


The blue line indicates the base score when using either an air or water cooling mechanism on both CPU and GPU. The red line indicates the performance increase when changing the CPU to a subzero cooling mechanism (e.g.: phase-change or liquid helium). The variation between cards can indicate the level of CPU dependant a video card really is, or, to what extend your CPU is bottlenecking the performance of your video card. In general, the more high-end VGA cards will have the most benefit from a higher clocked CPU.

The green line indicates the performance increase going from an air/water cooled configuration to a subzero cooled configuration or, to make it more clear, from FULL air/water to FULL subzero and NOT just changing the GPU configuration to subzero. The reason is quite simple: we assume that whoever is testing (or rather: submitting scores with) the GPU under LN2 will also have the CPU subzero to maximize the scores. Very important assumption!

The last step is to figure out whether it’s interesting to cool the GPU with liquid nitrogen or not. To find an answer to this question, we measure the difference between two test environments: only CPU subzero and both CPU/GPU subzero. The difference between both indicates the performance increase when changing the cooling method of the GPU.


data.pngWithout too much effort, I believe it’s fairly easy to conclude that the Nvidia Geforce GTX 275 is quite an interesting choice to use LN2 with. On the bottom of the list, we see that the Radeon 5970 is not really that interesting since the performance difference between CPU only and cpu/gpu subzero is a mere 3%. A very big side note to make here is that this doesn’t mean the card wouldn’t be able to scale more. It’s important to understand that since the data has been pulled from the HWBOT database, we make the assumption that if the card doesn’t scale properly with subzero cooling, people will not attempt to try LN2. This means that as we don’t have any control over the behavior of the overclocker, it’s not possible to determine whether not switching to LN2 means there’s no performance scaling or there’s no interest because it’s too difficult or too dangerous.

In general, however, we can see that the difference between two high-end models (relative to their release date) is most definitely noticeable. Something else noteworthy is the fact that most of the cards in the upper region of the ‘worthiness’-list are Nvidia, which would indicate that (in general) Nvidia scales better with cold. Therefore, I’ve also compiled an Ati versus Nvidia chart averaging the Ati and Nvidia performance results split up per temperature category.


Although a load of assumptions and side notes have to be made to this chart, it seems to be that in general Nvidia’s latest graphics cards scale better with temperature than Ati’s counter-products do.

If I had the time and the financial resources, I could perform similar research using different benchmarks and different hardware and even expand the cooling resources or vary platforms. Maybe we could even integrate this kind of information on the information pages, but sadly enough the current situation doesn’t allow us to do this. But, to end on a positive note, it does seem that the HWBOT database has a lot more to offer than just the raw scores, but can give us a better insight on the behavior of processors, video cards and recently also mainboards and memory when being pushed to the absolute limite in terms of operating frequency and operating temperature.


Belgium Massman says:

FYI - I used the 3DMark06 benchmark because it's by far the most popular 3D benchmark at HWBOT.

Romania Monstru says:!!!! :)

United Kingdom borandi says:

Nice comparisons. How about comparison including -ºC GPU/ ºC CPU, for people who use LN2 on the GPU and not the CPU? Also, the -ºC includes SS, Dice, etc. How about comparisons with those too? There may not be enough scores for that, but even one or two would be interesting.

Belgium Massman says:

Well, there's a reason why I made the assumption that people don't go LN2 on GPU without proper cooling on the CPU: CPU air/water + GPU subzero = 133 3DMark06 results CPU air/water + GPU subzero = 718 overall results To compare: CPU SS + GPU subzero = 122 3DMark06 CPU SS + GPU subzero = 653 overall CPU DICE + GPU subzero = 174 3DMark06 CPU DICE + GPU subzero = 917 overall

Germany der8auer says:

Really nice article!

jabski says:

nice article :)

Belgium Massman says:

I'm currently compiling a couple of charts in which I use the HWBoints as measurement tool. If there's any interest in those, I'll post them here as well :-)

Slovenia tiborrr says:

Thanks for posting this Massman, great reading.

United Kingdom r1ch says:

Pieter, is it worth adding another angle to this - I'm thinking hwboint gain from going sub-zero. Yes, it may only be a 10% gain in terms of score, but as you're overtaking so many people with that 10%, it could be a 100% difference in points - 30 points in air/water or 60 on dice/ln2?

Austria Turrican says:

i love those charts. :D

Russian Federation Antinomy says:

Pieter, how about the hwboints distribution between CPU and GPU? And with highlighting popular families. Like - the boints before HD2000 / 8000 and after comparison. The same for Core2 and pre-Core. And showing how much percents a certain CPU - Q6600, E8400, E8600 has from the overall points.

United States steponz says:

You have to love the powers of query... Working on a reporting engine at all??? Would be cool to see graphs like this, but in an automated fasion.

Eeky NoX says:

Really interesting work. Tough job there ! Congratz for those highlightnings Massman ;)

Germany BenchBros says:

Very interesting. Great work and thanks for these information *thumbs up*

Belgium Massman says:

r1ch said: Pieter, is it worth adding another angle to this - I'm thinking hwboint gain from going sub-zero.

Yes, it may only be a 10% gain in terms of score, but as you're overtaking so many people with that 10%, it could be a 100% difference in points - 30 points in air/water or 60 on dice/ln2?

Yeah, was thinking about that too, but I found it too 'dangerous' to use the hwboints as measurement tool to determine 'LN2 worthiness' with all those people screaming that benching is not only about hwbot. Of course it not, but just trying not to send out the wrong message.

I did run the query for with the points as measurement tool and I must say that it looks interesting. so:

- independent variable = temperature (~ article)
- dependent variable = Average global points (= global + wr points)
- using all benchmarks, not only 3DM06
- still single card only

Just one chart, the 'GPU LN2 worthiness chart', which is showing the difference between (CPU subero + GPU air water) and (CPU + GPU subzero).

So, for understanding this graph you need to understand that the line doesn't represent the amount of points itself, but the gain in points switching from CPU-only to full system subzero. So, although the increase may be spectacular it doesn't mean that you'll get the most points for it. The best way to interpretend this graph would be: "if I only have videocard X, should I push it under LN2 or not".

Maybe this is a bit more helpful chart, although it kinda says to go for the GTX480 if you want decent points. Probably, this is a bit biased because of the very small amount of submissions of the GTX480 at this moment, so the average is biased by outliers, but you get the point.

It might be cool to do this for every benchmark, videocard and also hardware points seperatelly. So, if you submit a score, the engine could tell you how much points there are to gain from switching cooling method on the GPU.

Antinomy said: Pieter, how about the hwboints distribution between CPU and GPU? And with highlighting popular families. Like - the boints before HD2000 / 8000 and after comparison. The same for Core2 and pre-Core. And showing how much percents a certain CPU - Q6600, E8400, E8600 has from the overall points.

Hm, may be a lot trickier than you think, because there's a difference between hardware points, global points and wr points. And what do you mean by 'before'? Also, the percentage of what exactly? Hardware points? Or all points?

I ran a quick query to see how much points are distributed in total:

- WR points = 2470
- Global points = 583713.2
- Hardware points = 473653.8
- Total = 583713.2 ~ 'half a million points' :D

steponz said: You have to love the powers of query... Working on a reporting engine at all???
Would be cool to see graphs like this, but in an automated fasion.

A reporting engine?

Anyway - Yes, it's perfectly possible to make this kind of automated charts, but as I mentioned in the article: there's no time and certainly no finances to provide all this.

Russian Federation Antinomy says:

I mean only the hardware points obviously - rather had to make globals on 6600GT for example :D By before I mean - the summarized HWboints for CPU or Video before that family and by after - the summarizing by family for example. Like this - 1. The old ones 2. The 8800/9800/HD2000/HD3000 3. The GT200 / HD 4000 4. Fermi / HD5000 And maybe make a chart to highlight the percentage that popular models take. It's not a secret that 8800GT/GTX take in sum much more boints than 8400GS/8500GT. So it would be nice to see the fatties stand-alone. Did you get my idea? Because of the age it's much more interesting to look only HW points, not global and 3D WR. I think the sum of points of about 10 cards will be bigger than all CPU results at least at pre-Core age.

Belgium Massman says:

The percentages and summarizing per family will take too much time to do it properly, but I did ran the query to check the last sentence. Turns out, you are right: the 8 most competitive videocards gather roughly the same amount of hardware points as all the cpus combined, excluding all the Intel Core series. The 5 most competitive CPUs also get the same amount of points as all the pre-Core cpus combined.

Russian Federation Antinomy says:

So no old school benching for points - just for fun. This is what I was telling a number of times and got a proof on that. Thank you :)

Belgium Massman says:

Well ... yeah, if there's no competition, there aren't a lot of points awarded to those categories. Do note that 'old-skool' also means 'pre-hwbot'. If HWBOT was launched in 2000, I'm sure the Athlon XP categories would be very rewarding at the moment.

Russian Federation Antinomy says:

Despite the fact they weren't yet developed? :D I mean the A-XP and year 2000. But I got your idea. Also it's the fact that the popularity grows in time so it wouldn't be as popular anyway. The vga vs cpu points is easy to explain - not everyone takes the top CPU because it's easier to overclock a cheaper one (and the difference between them in one family is percents or one or two tens of percents). But not for video - less models first of all. And because the difference in one family can count in times more people prefer the top product.

Belgium Massman says:

I'm not too confident of jumping to conclusions about the VGA cards, because there are a couple of different scenarios: - Submit a result with hardware you already have - Submit a result with hardware you bought --- for hardware points --- for global points Especially the last scenario (in which people buy a card for more global points) is one that should determine the amount of competition in a category, since most of the time only 1 or 2 cards can compete at the very top. I notice that the 8800GTX, 4870X2 and 5870 are gathering most hardware points and I'm sure the global points have something to do with that.

Russian Federation Antinomy says:

Maybe - I didn't take this in case. You are right. :)

Slovakia A2C says:

Good review! Here is somthing what solve our questions.

Please log in or register to comment.