Overclock.net banner

2741 - 2760 of 2781 Posts

·
Registered
Joined
·
150 Posts
Setup. Passive cooling only Airplex, external radiators 400 rpm e g silent mode. Delta T 2.9 water in vs ambient.

26 fans in Airplex (2x180, 24*140) and 120mm on the 3x360 + 1x240 (who has push pull). Silverstone on 180mm and Noctua ippc 3000 rpm on all other.

Loop order:

Reservoir->pumps->360/240 radiators->[water in pc] cpu->gpu (+mp5 parallel) [water out pc]->airplex->water filter->Reservoir
Ambient is a sensor on the floor as the top left temp is heated up a bit by the radiator air.

One note, with very low flow of say 50l/h the airplex alone cools the water more than 7 degrees (!) but the water temp (and delta T) of course goes up much higher than with close to 200 l/h flow so the gains are not there anyway except for if you want to have it really quiet but are okay with temps a bit higher.

In performance mode (pumps 75%, flow 200l/h and fans at 2000 rpm it is quite loud but delta T is very low). The room gets heated up also, hence a large evacuation fan or a small crack on the window is required on long sessions in order to keep ambient on a flat level. (the room is 6 sqm only)
Fans at 3000 rpm gives no real benefits except for more dBa. I have not seen more than 0.1c-0.2c decrease of water temp compared to 2000 rpm.

2478486


On standard load (gaming) with fans high and pumps 100%:

2478530
 

·
Registered
Joined
·
123 Posts
Some crazy builds...
Holley molley.
Talk about overkill. But I was pleasantly surprised that you don’t hit diminishing returns yet with that many radiators. I was under the impression that the issue with the temp was due to limited heat transfer from cpu die to the ihs which for 10980xe isn’t that bad because it is soldered but the size of the die (or rather the die density) is still limiting factor nonetheless and then 2nd weak point is from ihs to the cpu block.

I have 4 radiators, 2 x HWLabs 560GTR, 1 x EK PE360 and 1 x HWLabs 280 GTS all with Noctua ippc 3000 however my temp is still ****.

I’m building new custom case but I design them just to utilize the 2 x HWLabs 560GTR since I didn’t see improvement by using 4 rads. Probably will helps at lower rpm but even if I run all of the fans at 3000 rpm on all 4 of them, the temp spike is what caused the temp to reach limit not during constant load. The radiator fans is reacting fast enough since I control them using cpu temp which produces annoying noise compared with if I use water temp to control them. So I suspect my flow rate is the limiting factors I need to add more pump.

I also planning to replace the cpu block from velocity to either TechN or Optimus. TechN received better performance review with less restriction but **** build quality, anyway they are cheaper by a lot also they are in stock).

You said you are using quad d5 pump? How do you chain them in the loop? If I see correctly you just put them all in series before the whole loop? Have you tried to put them each after restriction? For example pump > cpu > pump > gpu > pump > rad. Something like what I have in mind for my custom case build. I think there is risk of cavitation since the second pump will be slightly starved due to the previous restriction but then I think in a closed loop with 1 pump only will also experience the same.
 

·
Old and Decrepit
Joined
·
1,877 Posts
45 min CPUz stresstest 100% usage. Again gaming never over 65c on these clocks.
I lowered max core to 4.9 as on 5.1 I need 1.365 which is a bit too much in my taste thinking about it. 5.0 is 1.351 so the step to 4.9 as in the pictures feels better.

Edit: I think I found the misunderstanding. I said "This is my 24/7 setup and when gaming the hottest core does not even reach 68c. Cpuz stresstest and hottest core 84c. This is with ambient 24c and water 25c "
Not that hottest core in a stress test is 68. Then I need to have open window and lower ambient of course.
My radiators are 3360 airplex gigant and 4x360, 1x240. Quad d5 pumps.
What are you stress testing with?
Thats some serious restrictions you have going on somewhere or have those D5s dialed way back. Im running 3 and if I put them all back to minimum Im still cranking out 400L/hr or better.

Try this stress test Blender full run (This is AVX)...18 cores 5GHz ambient 23C
This was my best run and on the edge of stable. It will crash eventually at that voltage. Once I pass 4.9 stability takes 1.32 or more depending on what else is OCd. If I have my mesh up and memory up its 1.34 and 1.35 to get 5.1


2478564
 

·
Registered
Joined
·
150 Posts
I ran cpuz and the stress test in it. Aida64 cpu stress gives less heat (without fpu).
Yes I do not know where the restrictions are. At 75% on pumps still only 190l/h.

I read it is diminishing return over 1 gpm but as I cannot reach it I cannot try.
Also I moved it out of the case as it was impossible to fit and temps was much worse as well as maintenance issues and more potential possible leak points.

Equipment water pass through:

Aqua computer filter
Aquacomputer highflow meter
Aquacomputer next highflow meter
Aquacool 10/13 fittings mix of 45/90/0
Softtube 10/13, 3-4m total or so
Techn 2066 block
Bykski 3090 block
Mp5 block in parallel (have serial top I will try)
3 alphacool industrial 360 rads
1 alphacool industrial 240 xflow rad
1 airplex 3360 rad
4 Alphacool quick couplings for high flowrate
Aquacomputer aqualis reservoir with fountain effect
2 d5 aquacomputer pumps
2 d5 next aquacomputer pumps

Where should I start to look ? (Bypass tests I guess)

Easiest to check is quick fittings, airplex and filter due to placement of them. I will report back but if you have any advice let me know.

Pressure drop on techn was not given by them but max pressure 1.5 bar their support told me. I asked bykski but no reply, did anyone test this on their gpu block ? I can test it as I have a spare pressure meter but I rather not tear down loop and **** right now.

Regarding diminishing return it is reached closer to 0 deg delta T in my view. At least in theory you should be close to infinity rad size then. In practical terms maybe at 0.5c or 0.7c or so is my guess on where it does not really impact much any longer.
 

·
Registered
Joined
·
123 Posts
The difference between 68 SP and 93 SP is enormous.
This is the VID of my new chip with 93 SP.
Will not OC it until I dial in the memory but this definitely can go all cores 50.
Temp was much better and so far seems a bit stronger IMC too.
38-40 is AVX-512 VID and 41 - 48 is SSE and AVX-256 VID.

2478585
 

·
Registered
Joined
·
150 Posts
The difference between 68 SP and 93 SP is enormous.
This is the VID of my new chip with 93 SP.
Will not OC it until I dial in the memory but this definitely can go all cores 50.
Temp was much better and so far seems a bit stronger IMC too.
38-40 is AVX-512 VID and 41 - 48 is SSE and AVX-256 VID.

View attachment 2478585
Agree. 71 is my chip :(
 

·
Registered
Joined
·
123 Posts
Agree. 71 is my chip :(
I haven't finished tuning my memory but I just can't wait to see where it lands.
God I love this new chip.
The old one needs 1.3v for 4.8 and 1.195v isn't even stable at 4.7.
Thing is not only the VID is lower but it is more consistent across the cores so I don't need to bother with per core.
Even running VID so I can use adaptive at 4.8 is still viable in my book after struggling with my old chip.

Below is just cinebench run but it also linpack stable. Haven't tried prime yet.
It also runs cinebench at 1.165v (just few times quick run so probably will fail eventually and at 1.165v it is not linpack stable).

Oh and this is at stock cache. Haven’t get around to oc it.

2478618
 

·
Registered
Joined
·
123 Posts
Damn the memory channel ABC are great they can do 4000T1 however channel D is **** can’t do 4000T1.

So unless I run 3 channels I have to choose 4000T2 or 3800T1.

Anyway this is much better than previous chip which can do 3800T2 only.

Will play with odt see if it helps.
 

·
Old and Decrepit
Joined
·
1,877 Posts
I ran cpuz and the stress test in it. Aida64 cpu stress gives less heat (without fpu).
Yes I do not know where the restrictions are. At 75% on pumps still only 190l/h.

I read it is diminishing return over 1 gpm but as I cannot reach it I cannot try.
Also I moved it out of the case as it was impossible to fit and temps was much worse as well as maintenance issues and more potential possible leak points.

Equipment water pass through:

Aqua computer filter
Aquacomputer highflow meter
Aquacomputer next highflow meter
Aquacool 10/13 fittings mix of 45/90/0
Softtube 10/13, 3-4m total or so
Techn 2066 block
Bykski 3090 block
Mp5 block in parallel (have serial top I will try)
3 alphacool industrial 360 rads
1 alphacool industrial 240 xflow rad
1 airplex 3360 rad
4 Alphacool quick couplings for high flowrate
Aquacomputer aqualis reservoir with fountain effect
2 d5 aquacomputer pumps
2 d5 next aquacomputer pumps

Where should I start to look ? (Bypass tests I guess)

Easiest to check is quick fittings, airplex and filter due to placement of them. I will report back but if you have any advice let me know.

Pressure drop on techn was not given by them but max pressure 1.5 bar their support told me. I asked bykski but no reply, did anyone test this on their gpu block ? I can test it as I have a spare pressure meter but I rather not tear down loop and **** right now.

Regarding diminishing return it is reached closer to 0 deg delta T in my view. At least in theory you should be close to infinity rad size then. In practical terms maybe at 0.5c or 0.7c or so is my guess on where it does not really impact much any longer.
Im not wrapping my head around the flow rate. With 3 D5s more fittings than I can count, same filter, GTR 420 rad, GTS 360, 480XE, aquacomputer filter, aqauacomputer Flowmeter barrowCH Flowmeter (about to be replaced with a Next Flowmeter) Optimus foundation ( for now, same as the Sig V2 but I wanted to verify) amd two heatkiller blocks I maintain a much higher flow rate. I’ve not addded the MORA with 2 more pumps into the loop with quick disconnects yet but I’ll post back with the flow results when I do. I did see a significant reduction in lower flow rates with the Optimus blocks and with the EK magnitude although the magnitude was not as significant and has less restriction. My Guess is the TechN is probably about the same as the Optimus in restriction. Design is not far off other than the TechN doesn’t have polished cold plates. I can’t see it’s performance being much different. I asked for a sample to test and didn’t even get the courtesy of a response telling me to pi$$ off. Only reviews I see are all AMD so can’t even compare someone else’s numbers.

I was wanted to see apples to apples. You evidently have a significant advantage over me ATM in rad surface area.

The AIDA64 CPU stress test isn’t very stressful. I don’t even pass mid 50s on temps, CB23 is actually more of a load. The FPU will slam it good though, most likely to the point of throttling without an AVX512 offset. I had to put in an offset of 3 at 4.8 to stay out of the red with the FPU test and even then it pushes the hairy edge in mid to upper 90s. Run the blender complete Benchmark without offset (which is an AVX (not 512 FPI) and it’s a good medium without going to extremes.

Ao long as it runs to your liking for your use that’s the main thing. Just looking apples to apples chip comparisons which we can’t do running completely different settinga. I’m on all core and you are not and the two stress tests are worlds apart. AIDA64 is a good tool. I have it interacted into my aquaero which was why I bought it in the first place. I can get VRM temps from it as well as some other goodies. About the only testing tool I use on it is the memory and Cache benchmark as I find the CPU and all other stress tests combined too weak and adding FPU will certainly heat check it for you and is way beyond what I’ll ever use it for as I don’t crunch primes all day. I do however do a lot of rendering and it’s usually using both the CPU and GPUs . You have to have the Intel driver installed to get that to work in open CL but not for some others.
 

·
OC...D
Joined
·
28,109 Posts
Setup. Passive cooling only Airplex, external radiators 400 rpm e g silent mode. Delta T 2.9 water in vs ambient.

26 fans in Airplex (2x180, 24*140) and 120mm on the 3x360 + 1x240 (who has push pull). Silverstone on 180mm and Noctua ippc 3000 rpm on all other.

Loop order:

Reservoir->pumps->360/240 radiators->[water in pc] cpu->gpu (+mp5 parallel) [water out pc]->airplex->water filter->Reservoir
Ambient is a sensor on the floor as the top left temp is heated up a bit by the radiator air.

One note, with very low flow of say 50l/h the airplex alone cools the water more than 7 degrees (!) but the water temp (and delta T) of course goes up much higher than with close to 200 l/h flow so the gains are not there anyway except for if you want to have it really quiet but are okay with temps a bit higher.

In performance mode (pumps 75%, flow 200l/h and fans at 2000 rpm it is quite loud but delta T is very low). The room gets heated up also, hence a large evacuation fan or a small crack on the window is required on long sessions in order to keep ambient on a flat level. (the room is 6 sqm only)
Fans at 3000 rpm gives no real benefits except for more dBa. I have not seen more than 0.1c-0.2c decrease of water temp compared to 2000 rpm.



On standard load (gaming) with fans high and pumps 100%:

View attachment 2478530
yeah, that's why I went with the Gigant sans aquaero... a 180 mounted under the grill and a cardboard cutout in the bottom mount position make the unit incredibly efficient. enough to cool anything I've thrown at it. The Aq is mounted in a bay or slot. On the Aquaduct 720 I have, the controller is not in the air flow channel. That design of the Gigant rad system always bugged me.

2478827
2478828
2478829
2478830
 

·
Registered
Joined
·
123 Posts
Tweaking the ODT allows me to run 4000 T1 on all 4 channels lol.

However only 2 out of 4 my sticks capable or running at 4000 T1 (they are 2 tottaly different kits).

And only 1 out of 4 sticks can run at 4000 cl15.

So for the time being need to fall back to 4000 T2 cl16 until I find good kit at good price they are sold everywhere.
 

·
Registered
Core i9 10980XE - Rampage VI EE - GSkill Trident Z 32GB 4400Mhz CL19 - Asus Strix 3090 OC - AX1200i
Joined
·
58 Posts
Tweaking the ODT allows me to run 4000 T1 on all 4 channels lol.

However only 2 out of 4 my sticks capable or running at 4000 T1 (they are 2 tottaly different kits).

And only 1 out of 4 sticks can run at 4000 cl15.

So for the time being need to fall back to 4000 T2 cl16 until I find good kit at good price they are sold everywhere.
Are you able to run 4000 Cl16 on Default Cache Ratio ( 24 ) or overclocked ratio ( 30 or 32 ) ? , because when I was trying to get 4000C16 stable with my 10980XE what I observed was that it is fully stable ( including GSAT ) at default cache ratio ( 24 ). However once cache is overclocked to 30/32, it destabilizes RAM OC and I start to get bit flip errors on GSAT tests even though it boots fine @ 4000CL16.
 

·
OC...D
Joined
·
28,109 Posts
You may need to add more VCCIO and Vcache, but what you are seeing is pretty common as long as you are not letting the sticks get too hot. Also, try decreasing tREFI with the cache at 30 or 32 if you have tREFI at a value >2x what auto sets.
 

·
Registered
Joined
·
123 Posts
Are you able to run 4000 Cl16 on Default Cache Ratio ( 24 ) or overclocked ratio ( 30 or 32 ) ? , because when I was trying to get 4000C16 stable with my 10980XE what I observed was that it is fully stable ( including GSAT ) at default cache ratio ( 24 ). However once cache is overclocked to 30/32, it destabilizes RAM OC and I start to get bit flip errors on GSAT tests even though it boots fine @ 4000CL16.
Like what jpmboy said probably you need more vccio since it can be considered the voltage for the interface between memory and cache. Try to leave the vmesh at stock auto and dial the vccio first, afterwards lower the vmesh if possible.

I have not oc the mesh at all on this new chip so it is at 24x however based on my experience with my old chip it doesn’t matter.
 

·
Registered
Core i9 10980XE - Rampage VI EE - GSkill Trident Z 32GB 4400Mhz CL19 - Asus Strix 3090 OC - AX1200i
Joined
·
58 Posts
You may need to add more VCCIO and Vcache, but what you are seeing is pretty common as long as you are not letting the sticks get too hot. Also, try decreasing tREFI with the cache at 30 or 32 if you have tREFI at a value >2x what auto sets.
I did try going upwards on both VCCIO and VCache in various combinations till 1.25V but didn't help. Reducing TREFI did give more mileage but still was conditionally stable. Since It can run 3800C15 ( faster in latency than 4000C16) with [email protected] 30/32 and at extreme tight secondaries & tertiaries with stock VCCIO and VCache indicated the reason is something else. And as I mentioned in my earlier posts, it turned out to be the temperature of my b-die sticks. If I can keep the sticks cool it works great. One of the most finnicky memory sticks I have seen till date. Somehow the cache overclock is causing the DRAM signalling integrity to destabilize since increased cache speeds up both bandwidth/latency for a given memory frequency.

Btw I did also came across one of your posts on the Intel DDR4 stability threads that 4400C19 sticks didn't work that well for you on your x299 builds. Well i'm running with them so that can be something.

Like what jpmboy said probably you need more vccio since it can be considered the voltage for the interface between memory and cache. Try to leave the vmesh at stock auto and dial the vccio first, afterwards lower the vmesh if possible.

I have not oc the mesh at all on this new chip so it is at 24x however based on my experience with my old chip it doesn’t matter.
Tried all of them, in my case it's the memory sticks that are the culprit. They just don't like high temperatures for some reason. Just unlucky i guess.
 

·
Registered
Joined
·
123 Posts
I did try going upwards on both VCCIO and VCache in various combinations till 1.25V but didn't help. Reducing TREFI did give more mileage but still was conditionally stable. Since It can run 3800C15 ( faster in latency than 4000C16) with [email protected] 30/32 and at extreme tight secondaries & tertiaries with stock VCCIO and VCache indicated the reason is something else. And as I mentioned in my earlier posts, it turned out to be the temperature of my b-die sticks. If I can keep the sticks cool it works great. One of the most finnicky memory sticks I have seen till date. Somehow the cache overclock is causing the DRAM signalling integrity to destabilize since increased cache speeds up both bandwidth/latency for a given memory frequency.

Btw I did also came across one of your posts on the Intel DDR4 stability threads that 4400C19 sticks didn't work that well for you on your x299 builds. Well i'm running with them so that can be something.


Tried all of them, in my case it's the memory sticks that are the culprit. They just don't like high temperatures for some reason. Just unlucky i guess.
Have you tired playing with the odt?

I tell you my secret sauce, first set DRAM training to enable, that way it will drop the channel/stick combination which has issue. This will tell you which slot/stick combination has problem.

Then from there try to adjust odt. I would suggest to change odt park only on the channel which have issue. For example the mobo drop channel B so you only adjust channel B and leave the other channel at auto.

This method fixed my slightly worse channel D so I can run 4000T1 stable on all channels. Sad that I only have 2 sticks capable of 4000T1 though.

Note that this method is only helping to fix/balance mainboard trace issue and will not magically makes your stick better binned.
 

·
OC...D
Joined
·
28,109 Posts
I did try going upwards on both VCCIO and VCache in various combinations till 1.25V but didn't help. Reducing TREFI did give more mileage but still was conditionally stable. Since It can run 3800C15 ( faster in latency than 4000C16) with [email protected] 30/32 and at extreme tight secondaries & tertiaries with stock VCCIO and VCache indicated the reason is something else. And as I mentioned in my earlier posts, it turned out to be the temperature of my b-die sticks. If I can keep the sticks cool it works great. One of the most finnicky memory sticks I have seen till date. Somehow the cache overclock is causing the DRAM signalling integrity to destabilize since increased cache speeds up both bandwidth/latency for a given memory frequency.

Btw I did also came across one of your posts on the Intel DDR4 stability threads that 4400C19 sticks didn't work that well for you on your x299 builds. Well i'm running with them so that can be something.


Tried all of them, in my case it's the memory sticks that are the culprit. They just don't like high temperatures for some reason. Just unlucky i guess.
Yeah, 4400c19s (which I sold) and even 4800c18 royals (running 4400c15 on a R10A/8086K rig right now) did not do all that great on x299 with either my 10980XE or 7980XE (or 9900X for that matter) on 3 different boards . I absolutely agree... 3800c15 and even 3733c15 with tight secondaries is really quick and low latency. Bandwidth via AID64 is only one measure of ram performance. I currently run 3600c15 for 64GB with really tweaked seconds and thirds - no complaints when compared to 4000c16 with 32GB on either rig. On the 8 slot boards, completing the ram channels fully gives a bit of benefit over leaving B empty at any frequency with a T-top layout.
Overclocking the cache and IMC will always lower the ceiling for ram frequency. That is why when you see WRs being set for ram frequency... you'll notice that the CPU is commonly underclocked. :)
 

·
Registered
Core i9 10980XE - Rampage VI EE - GSkill Trident Z 32GB 4400Mhz CL19 - Asus Strix 3090 OC - AX1200i
Joined
·
58 Posts
Have you tired playing with the odt?
I haven't tried tweaking the DRAM ODT settings yet. Any baseline values to start with ? Will give it a shot today.
 
2741 - 2760 of 2781 Posts
Top