360mm good quality AIO vs custom loop for ~250-300W CPU

[User5566]

I currently have what I consider a pretty good AIO, a corsair i150 pro 360. I bought it 4 years ago for my Ryzen 7 3700x (yes I know an overkill, but above all I wanted a quiet pc and it fulfilled that need marvellously).

However, needs change, and I've recently explored overclocking a ryzen 9 5950x at various powers up to 250W and recently I've upgraded to am5 ryzen 9 7950x which if overclocked can be pushed over 300W I believe. (BTW I use arctic mx-6 paste.)

When oc'ing 5950x I noticed a kind of a "brick wall" somewhere around 230W with my aio (once it reaches steady state). Until that power level temperature scales with increases in power, but definitely below a 1 to 1 ratio. For example at 225W I could keep my cpu at 83C(water temp around 36C, room temp about 21C). At 235W it very quickly reaches towards 90C. The water temperature is the same. So I have to assume the bottleneck in heat transfer is somewhere between the die, the ihs, the cold plate and the fin stack. Not the radiator.

Ryzen 9 7950x has a smaller ihs and I suspect this might be a reason why I see this very quick increase to even 95C at even lower power of 170W (no OC).

So I started researching custom loops (just for the cpu for now). And there is very conflicting info out there. There are videos that show 1C of difference at 230~300W power levels. And 15C difference at 350W(although on Intel).

Is it really possible to achieve 15C extra difference in temperatures in a custom loop at around 350W considering the radiator is essentially the same size and tubing thickness impact is negligible(as demonstrated elsewhere on this forum)?

The only real difference IMO between a good AIO and a custom loop seems to be water amount being pushed through the fin stack (due to much more powerful pump) and the height of the fin stack on the cold plate. Please note I'm not that interested in a brief advantage a custom loop can give you due to higher water volume available, but a steady state operation.

So in summary, has anyone replaced a 360mm AIO with a custom loop on a 350W CPU? What happened to your temps? What were the water temps before and after?

I have to admit I'm very surprised the water never gets hotter than 36.4C while the cores are 95C,but perhaps that's just the heat transfer across the ihs? Idk.

 

[Mr.Scott]

I will offer you a general answer. You get what you pay for in the WC world. AIO's are cheap mass produced items.
Everything is better in a custom system quality wise.

 

[User5566]

I will offer you a general answer. You get what you pay for in the WC world. AIO's are cheap mass produced items.
Everything is better in a custom system quality wise.

I don't think anyone can disagree with that, but if the other factors are causing the bottleneck having better components doesn't necessarily mean better performance.

So I'm hoping for some numbers

In theory there should be barely any difference.

What do I mean? Well, take the whole stack. The silicone, the indium solder, the ihs, the cold plate and it's fins(for now no thermal paste) .

(at 240W)
95C at the core, 84C at the top of the chip, 70C right beneath the ihs plate, 63C on the top of the cpu, then let's go for another 2.5mm of copper cold plate(including alongside fins) and we get ~59C contacting water.

(at 370W)
95C at the core, 78C at top of the silicone, 56C right beneath the ihs plate, ~46C on the top of the cpu, same 2.5mm of copper cold plate and fins and we get 40C contacting water.

So how much water at 36C do we need to take away 240W at 59C and 370W at 40C?

Not much. Barely anything.

From my "back of the napkin" calculation I get 9L/h and 82L/H. One d5 pump I found shows maximum flow of 1500L/h so I think the pump nor the tubing, not flow is the limiting factor.

The worst bottlenecks are I think (in the order of impact) :
- thermal paste/contact unevenness etc
- indium solder
- ihs itself

But the above is just theory. I'd love to hear real world results.

Even if you didn't have an aio, but you do have a ryzen 9 7950x with a custom loop. I'd love to hear what temps and scores you get in Cinebench r23 with your custom loop.

 

[EarthDog]

The worst bottlenecks are I think (in the order of impact) :

You forgot the silicon (silicone is for implants, among other things). The biggest factor is the density of the die itself...the smaller processes with more transistors in the same space the heat doesn't get out as easily. But a properly built 360mm custom loop is can be a few C better than your average 360 aio at the higher wattages. What that nets you... nit terribly much.

Flow rates should be in the ballpark of 1 GPM according to testing a while back... but that depends on everything used. 82 lph seems quite low for removing that kind of heat.

7950x on a 420mm aio.... as soon as I can fire up that pc, I'll let you know about cb23.

 

[Woomack]

If you compare something with the same size radiator, then it doesn't matter much ... as long as it's not a budget series AIO. On 240-360mm radiator, most top series CPUs will run at close to the max Tj or will throttle.
Most AIO brands use pretty good cold plates (micro-channels and other things), and pumps are good enough for a single radiator. However, most pumps in AIOs won't live as long as those used in custom liquid cooling. We also don't know what coolant is used or if the manufacturing process was clean enough to keep the coolant in perfect condition for longer. We heard that some users had problems with AIOs after a couple of months of work, and checking what was inside, they found something gooey.

Better performance have (well-designed) custom loops. The difference depends on the CPU and GPU. Usually, they're not worth the price if you use only a CPU. If you also use GPU, then the final result is much better, but for a higher series GPU+CPU, it's recommended to use at least two 240/360mm radiators or a single 420 mm+.

Another thing is component quality. Brands that seem to deliver top-quality components are not as reliable as all assume. Think that I'm in the process of removing the EK block from a motherboard, as all four screws/pins broke while trying to unscrew them with original tools. I will have to drill holes that will destroy the block, but hopefully, I won't damage the motherboard in the process. It's a long story, but it's not possible to do that another way. I had problems with multiple pumps, leaking radiators, blocks, broken fittings, and other things. Really, the list is quite long. Once everything is well, then the custom loop is a lot of fun, and the results are great. When things start to break or require maintenance more often than expected then it's only pain, wasted time, and money. This is why AIOs are so popular. You (usually) get a higher performance than that of air coolers, and it's easy to install, while manufacturers give up to a 5-year warranty.

 

[BugFreak]

The biggest difference I've seen in AIO builds and custom is the radiator thickness. If using slim radiators I doubt there would be much difference but putting a decent radiator in will make a difference. Personally I see no point in running AIO cooling when good air cooling will do just as good. Why add the risk of water when a good Thermalright or Noctua air cooler will do it with no risk.

 

[User5566]

Perhaps my expectations are just too high.

When there is one core at almost 90C and another one on the same die that is at 36C (water is 33C at this moment) that seems very weird to me... These cores are a mm apart on the die.

On the pic below thee power is only about 80W as I'm testing my undervolts now. I start to think it is the silicon that is the biggest bottleneck. And the only other thing one can perhaps do is to try keeping these water temps low so a bigger radiator as it was said above.



Also, I wonder, if I have an all core workload like prime95 and all cores run at roughly the same effective frequency but one core is significantly hotter than others, is there some way I can "level it out?". It seems that one core hitting 90C is holding the rest back. At the same time I can't give it more undervolt in curve optimiser, because the system will be unstable. Is this what the individual per core frequency limit is for? To undervolt, but without making the frequency too high?

 

[BugFreak]

I would try reseating the block in this situation.

 

[EarthDog]

I start to think it is the silicon that is the biggest bottleneck

That's what I said in post 4, yep! More rad isn't going to help much if the problem is inherently in the die.

You see that often depends on the load. If it's single-threaded, it can work like that. Notice CCD2, the 'furthest' away from CCD1 is cool and not touched comparatively? Core 4/5 i'd imagine are next to each other, look at their max temp...

You can try reseating, but I don't see a gross problem (many cores running warm temps) that would help. I don't see anything wrong here.

Just sitting on the desktop with a Twitch stream running (watching) and about 10 other FF tabs, with my 13900k and 360 AIO, one core peaks at 74C, and the rest of the P cores are in the 50s with one exception in the low 60s. The ecores are all within a few C of each other. But clearly, one thread is getting more work than another for brief periods on both of our systems. Again, seems normal to me.

Also, I wonder, if I have an all core workload like prime95 and all cores run at roughly the same effective frequency but one core is significantly hotter than others, is there some way I can "level it out?". I

It's not going to work that way........ run any Multi-threaded test and see... they'll be within a couple of C (normal) when you're stress testing because the load is similar across all cores/threads (as opposed to your 'idle' image above where background tasks spin up a core here and there and part of the reason you see those temps).

Personally I see no point in running AIO cooling when good air cooling will do just as good.

Noise.
Longer to get to a steady state.
Comparing best to best (360 AIO to air), air is still a few C behind last I checked.
And the risk, while greater than zero, is minimal.

Plenty of reasons.


EDIT: I have R24 scores for 7950X.... 2,155 (MT) and 124 (ST). That's with DDR5-6000. R23 is around ~38k.

 

[BugFreak]

Noise.
Longer to get to a steady state.
Comparing best to best (360 AIO to air), air is still a few C behind last I checked.
And the risk, while greater than zero, is minimal.

Plenty of reasons.

I should probably rephrase that. I see no reason to use AIO instead of air. If I am going water I will go full custom like my current setup.

 

[EarthDog]

I hear you. Just saying there are plenty of reasons to choose an AIO... or Air.........or water. Each has its positives and drawbacks over the other. Outside of performance and cost, it's a personal preference, really.

Some are OK with spending 3-4x+ as much for custom water loop while increasing the chance of leaking over an AIO for better performance and (potentially) even more quiet. Some want better than air peformance out of a sealed black box. Some want air.

 

[freeagent]

I don't know if there is a right answer as to which is the best, you would think a custom loop would be the absolute best, but I am not entirely convinced of that. I think each solution has its place, and if done right, each solution could provide the results you are looking for.

 

[Woomack]

Expanding some thoughts ...

Performance-wise, I see it like:
- Small air cooling (up to 120mm tall or just something designed for SFF) is about the same as a single 120mm AIO or a custom loop.
- Regular 140-155 mm tall, single tower coolers are about the same as 240 mm AIO/custom loop
- Top dual-tower air coolers are better than 240 AIOs, but worse than anything with larger radiators (in this group can be included top single tower coolers like Noctua NH-U12A)
- 280 mm AIOs or custom loops are about the same as 360 mm, assuming the same radiator thickness - AIOs are almost only available with slim radiators
- The top performance gives custom loops ... but expanding the loop above some point doesn't give any visible improvements, so it's like pumping money into aesthetics only

Noise-wise, I see it like:
- AIOs always have quiet pumps - if not, then they're broken or low quality, but most are Asetek-based anyway, and it's pretty good stuff. For custom loops, most people pick higher than required performing pumps, which are buzzing, so it ends up on limiting their speed.
- Most pumps have limited to about 60°C safe liquid temperature. You have to think about many things building smaller loops for higher wattage components or pumps won't last long. It means a custom loop that still makes more noise than expected.
- There are no silent "out-of-the-box" coolers for higher-wattage CPUs. It's not possible, and marketing lies. Everything at 150W+ makes noise. Most air coolers and AIOs make similar noises when they use the same amount of fans. There are differences between fans, but most options are 1500-2000 RPM. You set fan curves manually, or everything is noisy. You can spend $400 on the latest ASUS Ruijin III 360 AIO and find out it's very noisy while left at auto settings running with any high series CPU. The same you can build 6-12 fan custom loop only to find out that the noise is adding and all these fans make significant noise while set to only 700-800 RPM.

Price-wise, I see it like:
- for CPU only, custom loops are way too expensive - good stuff may cost $1k
- most AIOs are overpriced nowadays - some options go up to $500 ... but can find good options at a reasonable price, like well-performing and quiet, for example DeepCool AIOs
- air coolers will make more noise, but some reasonable options are starting from $50 - check Scythe Fuma 3 or some Thermalright options

Maintenance-wise, I see it like:
- Custom loop - problems are all the time to keep it clean and perform as new ... unless you are lucky or built an expensive loop in sterile conditions. Even if everything works then you always want to improve something and then you find out that the loop require changes, or at least cleaning.
- AIOs don't need much, but cleaning radiators can be annoying as it's often hard to do that right when they're installed; pumps die more often than quality ones in custom loops
- Air coolers - are pretty much maintenance-free, and unless a fan dies, they last for very long; they require cleaning occasionally, but the access is usually easy.

There is one thing which is annoying because of PC case manufacturers. You may have a hard time installing fancy, large custom loop in most new and well-looking PC cases. The main reason is that most brands make as much space as AIOs need, so 240/280/360mm long and 25mm thick. Usually, there is only space for a 25mm thick rad above the motherboard. Even if the manual says there is more space then is often hard to install it because of tall VRM heatsinks or no space for CPU power cables. I had a hard time finding anything reasonable for a 45mm thick radiators earlier this year. Even those options with spots for 3-4 360mm radiators, in reality support 1-2 radiators max and in others only fans can be installed (like most Corsair cases). Cases for radiators with 140mm fans are also hard to find and most support 25mm thick radiators+25mm thick fans.

My answer to the main question on this topic:
If you want a cooler only for a CPU, get AIO. If you add a graphics card, get a custom loop, but expect it to cost much more ... like even 4-5x more, and probably you don't expect that while planning it in your head. Building a custom loop for CPU only will not give you much better results than AIO, and will cost much more. A single 360 mm radiator is not so much worse than anything custom with more radiators as its thermal capacity is already 300W+. You may invest an additional $300-500 into 3-5°C under load. For some users, it's worth it. Most simply have no idea it's just a waste of money and time, as the average performance will be similar. Modern CPUs are not throttling if you use about as good cooling as a CPU's TDP - the difference is mainly in a shorter time for a maximum boost clock.

 

[User5566]

Wow, @Woomack thank you for a very comprehensive response. I think it covers pretty much everything.

My case is Thermaltake core X71. Good you mentioned ability to put thick rads on top. I just checked. My current AIO is 52mm thick(27mm Rad and 25mm fans) and if the rad was same dimensions, but 40mm the fans would clear 2mm above vrm heatsink... And there would be ~15mm space before it hits ram sticks so I could just do it.

BTW, I do have two very power hungry GPUs that would benefit greatly from a custom loop (gigabyte rtx3090 24g), but I'm having hard time getting water blocks for them here in Poland. It seems for every other rtx3090 you can get a local waterblock for well under $100, but for the gigabyte' gaming oc I'd have to get them from Aliexpress for $150 each (to be honest it's the 1~3 month wait that makes this the worst rather than $50 extra).

So with this cpu added that would be very close to 1kw in total power in worst case scenario. If I find decently priced waterblocks for my GPUs here in the EU I might do it.

Coming back to one of the sub-topics. This is about one of the cores being significantly hotter than others in a multi core workload like Cinebench r23 or other parallel calculations.

'Its not going to work that way........ run any Multi-threaded test and see... they'll be within a couple of C (normal) when you're stress testing because the load is similar across all cores/threads (as opposed to your 'idle' image above where background tasks spin up a core here and there and part of the reason you see those temps).

I am seeing that temperature difference here. The temperature difference is not that extreme of course, it's more like most cores are around ~80C-85C, but one core (core 5) goes to 95C in small ffts in prime95. (This was with curve optimizer set to negative 13,10,20,rest on 25).

I tried changing core 5 to - 28. It got even hotter (and became unstable). I changed it to - 20 and... The temps became much more level. Now I have Core 7 and 14 at the highest temps but as the difference between the coolest/hottest was over 10C, now it's more like 5C.

So, interestingly it seems I found another reason not to push the undervolt as low as it goes other than instability.

 

[Woomack]

Wow, @Woomack thank you for a very comprehensive response. I think it covers pretty much everything.

My case is Thermaltake core X71. Good you mentioned ability to put thick rads on top. I just checked. My current AIO is 52mm thick(27mm Rad and 25mm fans) and if the rad was same dimensions, but 40mm the fans would clear 2mm above vrm heatsink... And there would be ~15mm space before it hits ram sticks so I could just do it.

BTW, I do have two very power hungry GPUs that would benefit greatly from a custom loop (gigabyte rtx3090 24g), but I'm having hard time getting water blocks for them here in Poland. It seems for every other rtx3090 you can get a local waterblock for well under $100, but for the gigabyte' gaming oc I'd have to get them from Aliexpress for $150 each (to be honest it's the 1~3 month wait that makes this the worst rather than $50 extra).

11942 Alphacool Eisblock Aurora Acryl GPX-N RTX 3090/3080 Gaming/Eagle with Backplate

The Alphacool Aurora Acryl GPX-N RTX 3080/3090 Gaming/Eagle Eisblock combines style with performance and extensive digital RGB lighting. With…

shop.alphacool.com

Delivery 11-12 days and under 100 EUR. Price with the backplate.

You can install 2x 360mm rads in this case. It should be enough.

 

[User5566]

11942 Alphacool Eisblock Aurora Acryl GPX-N RTX 3090/3080 Gaming/Eagle with Backplate

The Alphacool Aurora Acryl GPX-N RTX 3080/3090 Gaming/Eagle Eisblock combines style with performance and extensive digital RGB lighting. With…

shop.alphacool.com

Delivery 11-12 days and under 100 EUR. Price with the backplate.

You can install 2x 360mm rads in this case. It should be enough.

Man, this is nice I thought only bykski and barrow make blocks for my card (gigabyte rtx3090 24g gaming oc). I saw some used Alphacool Aurora blocks for 3090FE and I assumed Aurora is just for FE cards. Clearly I was wrong. Now I checked their compatibility list Aurora 11942 is compatible.

The link you gave is great, but even better, now knowing what to look for I found a local distributor too.

So now I have a lot more options Thanks.

 

[Woomack]

You can order it from the Alphacool store. Shipping to Poland takes 2-3 days (1 day, but their logistics take 1-2 more). In short, up to 2 weeks in total and should be cheaper than from local stores.

 

[EarthDog]

I am seeing that temperature difference here. The temperature difference is not that extreme of course, it's more like most cores are around ~80C-85C, but one core (core 5) goes to 95C in small ffts in prime95. (This was with curve optimizer set to negative 13,10,20,rest on 25).

I tried changing core 5 to - 28. It got even hotter (and became unstable). I changed it to - 20 and... The temps became much more level. Now I have Core 7 and 14 at the highest temps but as the difference between the coolest/hottest was over 10C, now it's more like 5C.

I haven't seen an AMD chip or Intel chip be that off at default settings with a multi-core load before. Curious...

I don't play with a ton, but do play with more than the average user so this is a shock to see ONE core so much different.

@Woomack - have you seen this before?

 

[Woomack]

It shouldn't be more than a few degrees. I have +/- 2°C on 7800X3D right now (the same in CPU and CPU+FPU AIDA64 tests). On Intel it's typically up to 10°C between the best and the worst core. On AMD it's usually less.
It can be wrong reading or something with the CPU. As it was mentioned already, CPU's die is too small to make so big temp difference on a single core because of cooler mount.

I would ask AMD directly if it's an issue for RMA or not. I guess it would be the fastest way as you waste a lot of time on tests and the CPU will be probably quickly replaced. CPUs are barely ever tested as it takes too much time, so both, AMD and Intel replace them without deeper diagnostics.

 

[EarthDog]

I would ask AMD directly if it's an issue for RMA or not.

Perfect. This is what I was getting at. A normally functioning chip shouldn't be that off under multi-threaded loads. I'd be interested to hear what AMD has to say about it.

I'd love to see the Hwinfo log file to see what it's doing to clock speeds and if it's only on that core or others. But do agree chasing this down is a waste of time compared to a phone call/email to AMD support and telling them about it.