Fractal Design Celsius S24 Cooler Review

The liquid CPU cooling market has typically been divided into low-cost/closed-loop and high-priced/open-loop configurations, but manufacturers have attempted to bring forward products that were supposed to combine the best of both worlds over the past few years. Buyers of factory-filled open loops have been able get all the convenience of a leak-free factory assembly plus the option to open it later to add components. Yet there was never a best-of-both-worlds compromise on price, because those factory-filled loops deployed premium-priced components.

Fractal Design presents potential solutions in its 2x 120mm Celsius S24 and 3x 120mm Celsius S36 kits. We got our hands on the two-fan version in time for a launch-day review.

Specifications

MORE: In Pictures: 20 Clever Liquid-Cooled PC Setups

The first thing you’ll notice in the specs is that the radiator size is just that: the size of the radiator. Unlike most pre-filled open loops, the Celsius S24 doesn’t have its pump mounted there. Instead, we find a slightly oversized version of the water-block mounted pump familiar to the closed-loop market.

The radiator and pump are still equipped with standard G1/4 fittings to allow reconfiguration, just like competing solutions. And while the Celsius S24 doesn’t include extra fittings, it does include all the hardware needed to fit it to Intel’s three most recent generations of square-ILM sockets, AMD’s legacy rectangular-pattern sockets, and the new differently-spaced AMD socket AM4 mount. AMD users will need to remove the motherboard’s clip bracket, of course.

The Intel bracket is factory installed using a twist-connect design with locking tabs. Noticing that LGA 2011, 2011-v3, and 1366 all use the same spacing, Fractal Design even included a set of holes on its CPU support plate for LGA 1366. LGA 2011 and 2011-v3 have an integrated support plate.

The AMD bracket also uses AMD’s stock support plate, though “stock” is a loose term given that multiple designs have emerged. The support plates of most legacy boards have used standardized screws for many years, though earlier designs often used alternative attachment methods where the focus was on the clip rather than the bracket that held it. Even today, some AM4 support brackets have a threaded collar that extends past the motherboard’s top surface, while others don’t.

Fractal Design’s workaround for differing threaded collar heights is to include a special set of standoffs with stops that go around the collar and bottom against the board. These are polished, so they won’t scratch through the board’s protective mask and short any circuits. If you’re still more concerned about touching your delicate motherboard than mounting height, you’re free to use the traditional hardware that’s also included.

The Celsius S24 is powered at the pump, yet its braided sleeve perfectly conceals any hint of the embedded cable that runs from the pump to a breakout header for its fans. Pump speed can be regulated either by the motherboard’s PWM signal or via an internal temperature-based controller, and the pump relays its speed control to the fans. The above close-up also shows two of the four G1/4 fittings in greater detail.

Fractal Design’s Dynamic X2 GP-12 PWM fans are rated at 500-2000 RPM and a maximum 32.2 decibel (A-weighted). Fractal Design’s literature doesn’t mention the implication of a maximum of 35.2 for the pair, since doubling the number of sound sources adds 3db to a measurement.

MORE: Best CPU Cooling

MORE: How To Choose A CPU Cooler

MORE: All Cooling Content

Create a new thread in the US Reviews comments forum about this subject
27 comments
    Your comment
  • ingtar33
    looking good.

    My only question is the fans. I've loved FD cases, and owned a number, but the "stock" fans that come with those cases have been basically mediocre. While these FD fans appear to be quiet I wonder if we wouldn't see better performance and noise out of something else.
    0
  • dstarr3
    Anonymous said:
    looking good.

    My only question is the fans. I've loved FD cases, and owned a number, but the "stock" fans that come with those cases have been basically mediocre. While these FD fans appear to be quiet I wonder if we wouldn't see better performance and noise out of something else.


    I suppose if FD must cut costs somewhere, fans are a pretty good place to do it, since they are easily replaceable.
    0
  • jcwbnimble
    Could this cooler's lower l/h design be giving it the advantage compared to these other expandable coolers? By that I mean, with a lower flow, the coolant has more time to release it's heat in the radiator before returning to the CPU block. These other units have much higher flow ratings, most likely because they are anticipating the addition of GPU loops, and thus push the coolant thru the radiator before it's had enough time to dissipate the stored heat.

    Plausible explanation for the superior cooling achieved with a lower flow rate?
    0
  • DougDigsData
    Just a note on "flow" rates and that higher/more is perceived as better. In thermodynamics there is an optimal rate/velocity to use. Too "fast" - not good, too "slow" - not good. For example in AC systems in our houses, the velocity of the airflow through the indoor coil is set within a range to ensure the most effective heat transmission. That could be the situation with this unit as well.
    1
  • ingtar33
    Anonymous said:
    Could this cooler's lower l/h design be giving it the advantage compared to these other expandable coolers? By that I mean, with a lower flow, the coolant has more time to release it's heat in the radiator before returning to the CPU block. These other units have much higher flow ratings, most likely because they are anticipating the addition of GPU loops, and thus push the coolant thru the radiator before it's had enough time to dissipate the stored heat.

    Plausible explanation for the superior cooling achieved with a lower flow rate?


    that's plausible, though it would mean the radiators of the other coolers are poor for dispelling heat.
    0
  • TbsToy
    You guys are out of your minds. You say this is the best cooler you have tested in two years and yet say it isn't expandable. It isn't designed for expandability. Not many people are interested in water cooling. W.C. is way more trouble than it is worth. Try reporting with decency instead of your screwy nonsense. Wake up. Toms hardware used to be a good tech site but has turned in to a nonsense marketing thing.
    Walt Prill
    -9
  • RomeoReject
    Noob question coming up:
    If this is open loop, does that mean it needs to be maintained like a custom liquid cooling solution (Drained, filled, etc)? Or is it closer to a CLC in the sense it's more "mount and forget"? What's the advantage it has over either solution?
    0
  • Crashman
    Anonymous said:
    You guys are out of your minds. You say this is the best cooler you have tested in two years and yet say it isn't expandable. It isn't designed for expandability. Not many people are interested in water cooling. W.C. is way more trouble than it is worth. Try reporting with decency instead of your screwy nonsense. Wake up. Toms hardware used to be a good tech site but has turned in to a nonsense marketing thing.
    Walt Prill
    LOL, we said it was the best performing cooler tested on the big cooler testing system, which it is. That means it beat two dozen other big coolers including several enormous air coolers and all compact liquid coolers up to and including the 2x 140mm models.

    There's no marketing gimmick in seeing better numbers and reporting that the numbers are the best. Something tells me you were probably once a good observer of these details but that your observations have more recently turned into a nonsense counter-marketing thing.

    Eric did a special cooling project using multiple radiators on a different system. It's a full open loop. Putting all of that hardware into this system would alter the test parameters used to compare the other coolers, invalidating the comparison. But you may be interested in the components he used.

    Anonymous said:
    Noob question coming up:
    If this is open loop, does that mean it needs to be maintained like a custom liquid cooling solution (Drained, filled, etc)? Or is it closer to a CLC in the sense it's more "mount and forget"? What's the advantage it has over either solution?
    It's basically a CLLC that's had G1/4 fittings added: As long as you don't open it, it retains the CLLC basics.
    1
  • the nerd 389
    It's interesting that you rate the acoustic efficiency of the H220-X to be almost the same as the H240-X2.

    That seems fundamentally wrong, considering the H240-X2 gives a delta T of 51C while producing 29.8 dBA, and the H220-X produces 34.8 dBA to achieve the same delta T.

    Clearly, the H240-X2 is substantially more acoustically efficient.

    This also seems to underrate the efficiency of the Fractal. It manages to achieve a lower delta T than any other cooler at max, and that means overcoming the TIM in that CPU. Each additional reduction in temperature requires exponentially more power and airflow to achieve. The more realistic comparison would be to compare cooling power at equal noise levels or comparing noise levels at equal cooling. Anything else will severely distort the data unless the non-linear relationships between fan speed/noise, fan speed/air flow, and air flow/thermal conductivity are accounted for.
    0
  • mpampis84
    Does it perform better than the Kelvin S24 though? Kelvin prices have dropped recently, so is there any point in waiting for Celsius, apart from the integrated fan hub?
    0
  • Crashman
    Anonymous said:
    It's interesting that you rate the acoustic efficiency of the H220-X to be almost the same as the H240-X2.

    That seems fundamentally wrong, considering the H240-X2 gives a delta T of 51C while producing 29.8 dBA, and the H220-X produces 34.8 dBA to achieve the same delta T.

    Clearly, the H240-X2 is substantially more acoustically efficient.

    This also seems to underrate the efficiency of the Fractal. It manages to achieve a lower delta T than any other cooler at max, and that means overcoming the TIM in that CPU. Each additional reduction in temperature requires exponentially more power and airflow to achieve. The more realistic comparison would be to compare cooling power at equal noise levels or comparing noise levels at equal cooling. Anything else will severely distort the data unless the non-linear relationships between fan speed/noise, fan speed/air flow, and air flow/thermal conductivity are accounted for.
    It's not as much a rating as a calculation, we just put in the ∆T and the noise level and Excel spits out a number using basic math. Look at the noise difference as a percent, and the temperature difference as a percent, and then it makes sense.

    Anonymous said:
    Does it perform better than the Kelvin S24 though? Kelvin prices have dropped recently, so is there any point in waiting for Celsius, apart from the integrated fan hub?
    I don't believe we've ever had the Kelvin S24 on this machine.
    0
  • the nerd 389
    Anonymous said:
    Anonymous said:
    It's interesting that you rate the acoustic efficiency of the H220-X to be almost the same as the H240-X2.

    That seems fundamentally wrong, considering the H240-X2 gives a delta T of 51C while producing 29.8 dBA, and the H220-X produces 34.8 dBA to achieve the same delta T.

    Clearly, the H240-X2 is substantially more acoustically efficient.

    This also seems to underrate the efficiency of the Fractal. It manages to achieve a lower delta T than any other cooler at max, and that means overcoming the TIM in that CPU. Each additional reduction in temperature requires exponentially more power and airflow to achieve. The more realistic comparison would be to compare cooling power at equal noise levels or comparing noise levels at equal cooling. Anything else will severely distort the data unless the non-linear relationships between fan speed/noise, fan speed/air flow, and air flow/thermal conductivity are accounted for.
    It's not as much a rating as a calculation, we just put in the ∆T and the noise level and Excel spits out a number using basic math. Look at the noise difference as a percent, and the temperature difference as a percent, and then it makes sense.


    I know how you get the numbers. That's not the issue at all, and it never has been.

    The issue is that the numbers are presented as acoustic efficiency, and simply do not reflect the relationship between cooling power and noise output in practice. The end result is that these numbers are often cited to say product A is more acoustically efficient than product B. The problem as I see it is that in many cases, product B will provide equal cooling at a lower noise level.

    This is entirely counter intuitive to many readers, and it takes some effort on our part to determine which product will actually give lower noise in any given application. I would greatly appreciate it if I could read a review and know what to expect out of a product. As it is, I have to hunt through the charts to find what the delta T and SPL values are, and then figure out the relationship between them on my own.
    0
  • Crashman
    Anonymous said:
    I know how you get the numbers. That's not the issue at all, and it never has been.

    The issue is that the numbers are presented as acoustic efficiency, and simply do not reflect the relationship between cooling power and noise output in practice. The end result is that these numbers are often cited to say product A is more acoustically efficient than product B. The problem as I see it is that in many cases, product B will provide equal cooling at a lower noise level.

    This is entirely counter intuitive to many readers, and it takes some effort on our part to determine which product will actually give lower noise in any given application. I would greatly appreciate it if I could read a review and know what to expect out of a product. As it is, I have to hunt through the charts to find what the delta T and SPL values are, and then figure out the relationship between them on my own.
    It's the actual measurement of how much cooling you get per unit of noise. Or how much noise you get per unit of cooling. Actually it's both of those. It's like measuring how many work units you get per unit of electricity...so yes...a system that produces twice the data while using the same energy is twice as efficient, same deal using acoustical energy instead of electrical energy...

    We get the same feedback occasionally when a system that uses 100W to produce 100 work units is rated less efficient than one that uses 150W to produce 200 work units. "The 100W unit [must] be more efficient because it uses less power". Nah, it's just more miserly.
    0
  • the nerd 389
    Anonymous said:
    Anonymous said:
    I know how you get the numbers. That's not the issue at all, and it never has been.

    The issue is that the numbers are presented as acoustic efficiency, and simply do not reflect the relationship between cooling power and noise output in practice. The end result is that these numbers are often cited to say product A is more acoustically efficient than product B. The problem as I see it is that in many cases, product B will provide equal cooling at a lower noise level.

    This is entirely counter intuitive to many readers, and it takes some effort on our part to determine which product will actually give lower noise in any given application. I would greatly appreciate it if I could read a review and know what to expect out of a product. As it is, I have to hunt through the charts to find what the delta T and SPL values are, and then figure out the relationship between them on my own.
    It's the actual measurement of how much cooling you get per unit of noise. Or how much noise you get per unit of cooling. Actually it's both of those. It's like measuring how many work units you get per unit of electricity...so yes...a system that produces twice the data while using the same energy is twice as efficient, same deal using acoustical energy instead of electrical energy...

    We get the same feedback occasionally when a system that uses 100W to produce 100 work units is rated less efficient than one that uses 150W to produce 200 work units. "The 100W unit [must] be more efficient because it uses less power". Nah, it's just more miserly.


    I understand the logic behind that. Unfortunately, acoustics rarely works like that. The units for, say, a motor's efficiency are based on two similar units, namely mechanical vs electrical watts. Since any decibel value is a unitless ratio, you can't simply use it like you would other units.

    Acoustic efficiency of a fan (and yes, there is an official definition) is very different. Because the noise level produced by a specific fan scales with 10*log( RPM^5), airflow scales with RPM, and heat transfer scales with e^(-k*cfm) you cannot simply look at the difference in cooling vs the difference in noise. I can't even begin to cover all of the places where the math simply does not work.

    The official definition of acoustic efficiency is actually a simple dB offset that's added into the fan noise prediction equations, and is defined in terms of RPM, not delta T. No where will you find a dB/watt unit, a dB/C unit, or a dB/RPM unit in the field of acoustics. The closest you actually see is dB (SPL) - dB (watts). That's a minus, not a division. For fans, this comes into play as a specific 50*log(RPM1/RPM2) term in the noise equations.

    Mind you, there's still the exponential decay involved with the amount of cooling a fin array can give you with increasing airflow. That's probably the most annoying factor to account for due to modern heatsinks inducing turbulence as a means of increasing heat transfer. I'm unfortunately not in a position that I could test that, but TH reviewers might be. I understand the time limitations you guys face, though. I don't know of an efficient way to measure that off the top of my head, so it probably isn't practical to implement a measurement method without further research.

    I'd be happy to link in some reference material for you to go over if you're interested. I have some documents that may be worth going through, as they outline the key parameters used to describe fans in terms of acoustics. They're far more concise and understandable than I can hope to be.

    Minor Update:
    I feel like it's worth mentioning that I wouldn't bother the reviewers at TH with this if I hadn't gone through and triple checked both the methods that I'm suggesting and the methods currently used in the reviews. I don't care to waste anyone's time on trivialities or even minor errors. I only bring this up because it presents significant issues when selecting products, and can easily lead consumers to purchase products that are not what they want/need.
    1
  • Crashman
    Anonymous said:
    Anonymous said:
    It's the actual measurement of how much cooling you get per unit of noise. Or how much noise you get per unit of cooling. Actually it's both of those. It's like measuring how many work units you get per unit of electricity...so yes...a system that produces twice the data while using the same energy is twice as efficient, same deal using acoustical energy instead of electrical energy...

    We get the same feedback occasionally when a system that uses 100W to produce 100 work units is rated less efficient than one that uses 150W to produce 200 work units. "The 100W unit [must] be more efficient because it uses less power". Nah, it's just more miserly.


    I understand the logic behind that. Unfortunately, acoustics rarely works like that. The units for, say, a motor's efficiency are based on two similar units, namely mechanical vs electrical watts. Since any decibel value is a unitless ratio, you can't simply use it like you would other units.

    Acoustic efficiency of a fan (and yes, there is an official definition) is very different. Because the noise level produced by a specific fan scales with 10*log( RPM^5), airflow scales with RPM, and heat transfer scales with e^(-k*cfm) you cannot simply look at the difference in cooling vs the difference in noise. I can't even begin to cover all of the places where the math simply does not work.
    We've mentioned that the decibel scale is logarithmic and even mention in this case that twice the sound sources will typically increase the noise by only 3db (in other reviews we talk about 10db being twice the perceived loudness). Would you like to point out the formula that Excel could use, that would treat sounds as a logarithmic scale and heat as a standard scale?
    0
  • the nerd 389
    Mentioning that it's logarithmic does not mean you can then ignore the fact that it is, or that it's not directly (or even closely) proportional to things like delta-T.

    The simplest way to convert dB to a linear scale is 10^(dB/10) if you're interested in energy content, and 10^(dB/20) if you're interested in some physical quantity like pressure. This still won't correct for the fifth order relationship with RPM or the exponential relationship between that and delta T. There's also the constant k that needs to be taken into account, and that varies based on the construction of a cooler. I'll have to sit down with the math to be sure, but there should be a way to derive it from either two or, preferably, three measurements of delta T and the corresponding SPL.

    I'll see if I can put a formula together. I usually use either C# or VBA to make that sort of operation legible and easy to work with when dealing with this sort of thing at work. Otherwise, the formulas get rather cumbersome very quickly. If that isn't acceptable, I can write the entire thing out in a single formula.

    If you'd like, I can also throw in a term to correct for the thermal interface material used in Intel CPUs. Currently, I can only make a reasonable estimate of the i7-7700k, and that's +/-20%. You guys probably have enough data to derive it with much greater accuracy than I do. You also probably have enough data to get the numbers for other processors and product lines.

    I just have one question, though. Do you want the formula/code to be backwards compatible with the data you've gathered in previous reviews? If so, it may take some time to put together, even for me.
    0
  • Crashman
    Anonymous said:
    Mentioning that it's logarithmic does not mean you can then ignore the fact that it is, or that it's not directly (or even closely) proportional to things like delta-T.

    The simplest way to convert dB to a linear scale is 10^(dB/10) if you're interested in energy content, and 10^(dB/20) if you're interested in some physical quantity like pressure. This still won't correct for the fifth order relationship with RPM or the exponential relationship between that and delta T. There's also the constant k that needs to be taken into account, and that varies based on the construction of a cooler. I'll have to sit down with the math to be sure, but there should be a way to derive it from either two or, preferably, three measurements of delta T and the corresponding SPL.

    I'll see if I can put a formula together. I usually use either C# or VBA to make that sort of operation legible and easy to work with when dealing with this sort of thing at work. Otherwise, the formulas get rather cumbersome very quickly. If that isn't acceptable, I can write the entire thing out in a single formula.

    If you'd like, I can also throw in a term to correct for the thermal interface material used in Intel CPUs. Currently, I can only make a reasonable estimate of the i7-7700k, and that's +/-20%. You guys probably have enough data to derive it with much greater accuracy than I do. You also probably have enough data to get the numbers for other processors and product lines.

    I just have one question, though. Do you want the formula/code to be backwards compatible with the data you've gathered in previous reviews? If so, it may take some time to put together, even for me.
    Well yeh, I'd like to drop the formula into the existing spreadsheet if possible, since we have a couple dozen coolers already in it.

    The original sheet takes two averages for the two measurements (high and low) of all coolers, then divides that result by the single result to generate a percent scale inverse to the temperature (ie, 12% less heat than the group average of 100% gives it a rating of 112%). The division works the other way in the noise measurement, because that number is going on the bottom of the cooling-to-noise equation (ie, 12% more noise than average is 112%). And then the temperature % is divided by the noise % to get a cooling-to-noise ratio without the logorithmic scaling on the noise ratio. And then the average is zero'd out by subtracting 1 from the result and putting the -1 data on the chart. So, I think we'd need to go up to the middle, where the individual noise measurement is divided by the average for the group, in order to add the logorithmic to linear scale correction.
    0
  • the nerd 389
    I was thinking of calculating the rating based on the two noise and temperature measurements, rather than the average of the two or the average of the group. I might account for the RPM as well simply to ensure the consistency of the results.

    If I do that, I should be able to make the entire thing independent of the other coolers in the review. That's one less issue for the reviewers to worry about, no?
    0
  • Crashman
    Anonymous said:
    I was thinking of calculating the rating based on the two noise and temperature measurements, rather than the average of the two or the average of the group. I might account for the RPM as well simply to ensure the consistency of the results.

    If I do that, I should be able to make the entire thing independent of the other coolers in the review. That's one less issue for the reviewers to worry about, no?
    TBH, the only thing I can do with any degree of accuracy is to base gains or losses on the average. Otherwise I'm left doing calculations based on 100% efficiency being 0° gained at 0 decibels, and zero doesn't work so well in calculations :D
    0
  • Olle P
    Anonymous said:
    Could this cooler's lower l/h design be giving it the advantage compared to these other expandable coolers? By that I mean, with a lower flow, the coolant has more time to release it's heat in the radiator before returning to the CPU block.
    Low flow is detrimental to good cooling.
    You're correct that the coolant has more time to dump heat to the air, but:
    a) it will do so in a less efficient manner, since the average temperature gradient between coolant and air will be lower, and
    b) the coolant will also heat up more while staying longer at the cooling block.
    In reality there's a threshold flow level above which the cooling doesn't become significantly mor efficient, but below that level any change is clearly noticed.

    Anonymous said:
    The simplest way to convert dB to a linear scale ... still won't correct for the fifth order relationship with RPM or the exponential relationship between that and delta T. There's also the constant k that needs to be taken into account, and that varies based on the construction of a cooler.
    Neither changes in fan speed nor changes in construction are relevant for the data presented.
    The design is obviously fixed and measurements are done with the fans at full and half speed.

    My only gripe is that the presented noise level doesn't have the background noise removed. (Delta noise instead of absolute noise.)

    The way to calculate actual noise from the fan (N [dB(A)]) given background noise (B) and measured noise with fan (M) is N = 10*log(10^(M/10) - 10^(B/10))
    Note: N ≈ M if M > B+5
    0