new build 2.0
- Thread starter Psi*
- Start date
Psi*
Tech Monkey
I like the Indigo Extreme very much. I find it very easy to apply. I have never removed a heatsink in which it was used tho. I read someplace that it just peals off. It is rated as one of the best & for that extra couple of degrees for the duty cycle I run my machines thru ... almost all threads pegged for hours at a time, I am all over that.
I did back the OC down from 4.6+ GHz as I just could not get an extended run with Intel Burn In program.
I did back the OC down from 4.6+ GHz as I just could not get an extended run with Intel Burn In program.
Psi*
Tech Monkey
For me, it is coming down to the easiest design to clean the dirt out of some kind of fins. With a radiator a can of compressed air can be used to blow back thru the fins. Quick and dirty, literally.
With air cooled, and the box used in this latest WC system had a Noctua "tower" with push/pull fans. Getting those damn fans mounted back on that thing is nearly impossible and requires 3 hands the size of a 10 y/o which would still be challenging. But when you have the fans off, the dirt is packed into most of the fins and considering the angle are pretty impossible to figure out if you got them cleaned or not. So this is why I say that heatsink must be removed to make sure that you have it cleaned. Of course, it is often recommended on the net that TIM needs to replaced periodically so maybe this should be consider part of the territory. I don't see air cooling as a place that I would use Indigo Extreme assuming you agree with removing the heatsink for cleaning.
I really gave a lot of thought to using separable water fittings and having the whole radiator/fan assembly external and as a separate unit to the main computer case. I am still thinking that would be the best for routine convenient cleaning. Just thought that I would give the all in one another shot.
With air cooled, and the box used in this latest WC system had a Noctua "tower" with push/pull fans. Getting those damn fans mounted back on that thing is nearly impossible and requires 3 hands the size of a 10 y/o which would still be challenging. But when you have the fans off, the dirt is packed into most of the fins and considering the angle are pretty impossible to figure out if you got them cleaned or not. So this is why I say that heatsink must be removed to make sure that you have it cleaned. Of course, it is often recommended on the net that TIM needs to replaced periodically so maybe this should be consider part of the territory. I don't see air cooling as a place that I would use Indigo Extreme assuming you agree with removing the heatsink for cleaning.
I really gave a lot of thought to using separable water fittings and having the whole radiator/fan assembly external and as a separate unit to the main computer case. I am still thinking that would be the best for routine convenient cleaning. Just thought that I would give the all in one another shot.
Last edited:
Psi*
Tech Monkey
I did some experimentation today with IntelBurnTest.
The columns are (and they do not show up at all);
#threads memory GFlops temp.
HT on
12 max23170 56.5 50-60 ... this translates to 12 threads, maximum stress test with 23170 MB Ram, 56.5 GFlops, temperature via SpeedFan ranges between 50 to 60 deg C
12 vhi4096 86.6 60-70
12 high2048 74.2 55-65
12 std1024 70.5 55-60
6 std1024 87 55-60
6 high2048 93 55-62
6 vhi4096 94.5 55-65
6 max23589 94 58-64
HT on 2 instances ... each instance of IBT configured identically & run concurrently
6 std 48 60-65
6 high 39-49 60-70
6 vhi4096 49(42x1/1) 60-70
HT off
6 std1024 94.2 52-60
6 high2048 96.5 55-62
6 vhi4096 97.9 58-64
6 max22543 99.5 57-65
HT off 2 instances
6 std 92 55-62
6 high 95(48x1/1) 57-64
6 vhi 97(69x1) 55-62
I apologize for the poor formatting as I haven't copied/pasted a table into any forum before. A bit of a legend ... std is 1024 MB memory, "high" is 2048 MB, "vhi" is 4096 MB.
I am quite surprised about this. As you study this, it is clear that hyperthreading for IBT is not a good thing. In fact, IBT and HT seems to provide the worst possible performance in general and with a pretty high confidence. There are a few outlyers noted as 95(48x1/1) indicating that 4 runs had 95 GFlops but there were 48 GFlops in each instance of IBT.
The skeptic in me asks, "is this similar for other algorithms?" (especially the stuff that I run?) Of course no one here can answer the latter but myself.
Additionally the begging question is "what is going on even with just this 1 algorithm?" Almost a year ago in this thread SMP limited by memory controller I raise the question about the single memory controller in the i7 chips as being the neck of the funnel ... aka the restricting stage in the structure of I7s.
I have to admit that I do not understand near enough about this topic to offer an intelligent opinion. The supposed expert opinion in that thread is from the s/w developer of the s/w that I live & die in. The developers behind that s/w are literally all over the world & typically universities. Again the skeptic in me asks .... "are these students ... even wanna be PhD students, but with what focus?"
The columns are (and they do not show up at all);
#threads memory GFlops temp.
HT on
12 max23170 56.5 50-60 ... this translates to 12 threads, maximum stress test with 23170 MB Ram, 56.5 GFlops, temperature via SpeedFan ranges between 50 to 60 deg C
12 vhi4096 86.6 60-70
12 high2048 74.2 55-65
12 std1024 70.5 55-60
6 std1024 87 55-60
6 high2048 93 55-62
6 vhi4096 94.5 55-65
6 max23589 94 58-64
HT on 2 instances ... each instance of IBT configured identically & run concurrently
6 std 48 60-65
6 high 39-49 60-70
6 vhi4096 49(42x1/1) 60-70
HT off
6 std1024 94.2 52-60
6 high2048 96.5 55-62
6 vhi4096 97.9 58-64
6 max22543 99.5 57-65
HT off 2 instances
6 std 92 55-62
6 high 95(48x1/1) 57-64
6 vhi 97(69x1) 55-62
I apologize for the poor formatting as I haven't copied/pasted a table into any forum before. A bit of a legend ... std is 1024 MB memory, "high" is 2048 MB, "vhi" is 4096 MB.
I am quite surprised about this. As you study this, it is clear that hyperthreading for IBT is not a good thing. In fact, IBT and HT seems to provide the worst possible performance in general and with a pretty high confidence. There are a few outlyers noted as 95(48x1/1) indicating that 4 runs had 95 GFlops but there were 48 GFlops in each instance of IBT.
The skeptic in me asks, "is this similar for other algorithms?" (especially the stuff that I run?) Of course no one here can answer the latter but myself.
Additionally the begging question is "what is going on even with just this 1 algorithm?" Almost a year ago in this thread SMP limited by memory controller I raise the question about the single memory controller in the i7 chips as being the neck of the funnel ... aka the restricting stage in the structure of I7s.
I have to admit that I do not understand near enough about this topic to offer an intelligent opinion. The supposed expert opinion in that thread is from the s/w developer of the s/w that I live & die in. The developers behind that s/w are literally all over the world & typically universities. Again the skeptic in me asks .... "are these students ... even wanna be PhD students, but with what focus?"
Last edited:
Psi*
Tech Monkey
well who knew?!?!
I am finally back at work & with this new i7 990x loaded up with the s/w that I make my living by ... serious numbercrunching stuff. Due to recommendations by the s/w seller & the recent testing I have made on the 990X, I have HT off (still testing, but am now willing to try ... tho skeptical). Also commented by the s/w seller is that 6 cores may not perform significantly better than 4.
I am currently running a model that farms parts of itself out to each machine ... the i7 920 & the i7 990X. Each are OC-ed to ~4.5 GHz. The 920 is backed off due to higher temps and consequential crashes which indicates that the radiator needs to be blown out (I'll get to it this weekend).
Soooo, the interesting part is that there is no significant difference in thru put of these 2 machines! The 990X has 24 GB versus the 920s 12 GB and a initial pass at getting this thing done resulted in maxxing out the 920s 12 GB! Looking at the 990X, it hit 12+ GB so the model did not completely blow thru the roof of the 920.
So my 1st blush impression is to just get more RAM for the 920 which means I'll have 6 sticks of 2 GB Dominator RAM for sale soon to be replaced with 6 sticks of 4 GB RAM ... which now costs less than the 12 GB when I got them!
Wow!
I am finally back at work & with this new i7 990x loaded up with the s/w that I make my living by ... serious numbercrunching stuff. Due to recommendations by the s/w seller & the recent testing I have made on the 990X, I have HT off (still testing, but am now willing to try ... tho skeptical). Also commented by the s/w seller is that 6 cores may not perform significantly better than 4.
I am currently running a model that farms parts of itself out to each machine ... the i7 920 & the i7 990X. Each are OC-ed to ~4.5 GHz. The 920 is backed off due to higher temps and consequential crashes which indicates that the radiator needs to be blown out (I'll get to it this weekend).
Soooo, the interesting part is that there is no significant difference in thru put of these 2 machines! The 990X has 24 GB versus the 920s 12 GB and a initial pass at getting this thing done resulted in maxxing out the 920s 12 GB! Looking at the 990X, it hit 12+ GB so the model did not completely blow thru the roof of the 920.
So my 1st blush impression is to just get more RAM for the 920 which means I'll have 6 sticks of 2 GB Dominator RAM for sale soon to be replaced with 6 sticks of 4 GB RAM ... which now costs less than the 12 GB when I got them!
Wow!
Psi*
Tech Monkey
What I have learned this time ... general WCing building ideas
So somethings that I hope will stick with me from this build (990X) & the prevous (920) & listed in no particular order;
1) the plastic tubing I use is the cheapest 1/2" ID stuff that I can find. The logic is that it will need to be changed out in 1 to 2 years. That may not necessarily be true. But I buy an extra several feet of it anyway as it helps to lay things out & play around with bendy--ness ... how tight of bend is too tight? That leads to the next learned issue.
2) Anti-kink coils ... to use or not to use is the question. Added to with, "how much?" Answer to the 1st part is definitely "yes" use the coils as they definitely allow the tubing to be bent much tighter (to a point) than possilble without. The coils also provide a certain dress up to the system. I will admit to looking for black coils. Many coil colors exist in several different brands ... again I bought the cheapest black to contrast against my beige case (chagrin).
A special note about the coils, it is possible to put them on too close. This makes the tubing overly stiff & that translates to the tubing & fitting connection. This is what caused the problem I had with this (990X) build with the weepy fittings connections. The tubing + coils and my choice of using Tye-wraps on barb fittings == leaks. My solution was tighter & additional Tye-wraps to eliminate the leaks as I did not realize at that moment that I had wound excessive number of "winds" of the anti-kink coils on the tubing.
3) I am liking push-pull fans with "thick" radiators. The thick one is the Black Ice used in this build. Push-pull fans kind of guarantee good air flow thru the radiator. The 920 build used the thinner Switech radiator and I added a couple of the uber thick tfc triebwerk fans to it today. It did not change the idle temp but the under load temp dropped 10 degrees!!! Yaaay
4) In fact I highly recommend these tfc triebwerk fans for any radiator application. Yes they are hugely thick at 55 mm but that is really what it takes to force (or pull in this case) air thru the thicker radiators. These particular fans are low speed & therfore very quiet. Push-pull wouldn't really be necessary with these fans, but I already had the pushers ... so ... I kept 'em. I'll suggest a google search of "tfc triebwerk review" which will provide a number of reviews and youtubes.
5) fan filters ... doh! and duh! I just put these filters on every fan I own. I decided to try these versus foam. No solid reason as to why other than I wanted to try something between the foam and screen wire. I have pets & lets just say that animal fur becomes infused in all things involved in cooling a PC.
These particular filters come with 4 M4 X some unknown length screws ... perfect for the Black Ice but too big for the Switech radiator (in the 920 build). These tidbits would be great to be mentioned as I don't find them on any seller's web sites.
Perhaps this could be another unique to techgage review series? That would be evaluating fan filters. As an 'ngineer I think about all kinds of impossible tests ... air flow restriction being at the top of the list. Effectiveness being next, but I cannot think of a simple test for that ... maybe later.
There might be a few more thoughts, but this should be good for those thinking about a WC build.
So somethings that I hope will stick with me from this build (990X) & the prevous (920) & listed in no particular order;
1) the plastic tubing I use is the cheapest 1/2" ID stuff that I can find. The logic is that it will need to be changed out in 1 to 2 years. That may not necessarily be true. But I buy an extra several feet of it anyway as it helps to lay things out & play around with bendy--ness ... how tight of bend is too tight? That leads to the next learned issue.
2) Anti-kink coils ... to use or not to use is the question. Added to with, "how much?" Answer to the 1st part is definitely "yes" use the coils as they definitely allow the tubing to be bent much tighter (to a point) than possilble without. The coils also provide a certain dress up to the system. I will admit to looking for black coils. Many coil colors exist in several different brands ... again I bought the cheapest black to contrast against my beige case (chagrin).
A special note about the coils, it is possible to put them on too close. This makes the tubing overly stiff & that translates to the tubing & fitting connection. This is what caused the problem I had with this (990X) build with the weepy fittings connections. The tubing + coils and my choice of using Tye-wraps on barb fittings == leaks. My solution was tighter & additional Tye-wraps to eliminate the leaks as I did not realize at that moment that I had wound excessive number of "winds" of the anti-kink coils on the tubing.
3) I am liking push-pull fans with "thick" radiators. The thick one is the Black Ice used in this build. Push-pull fans kind of guarantee good air flow thru the radiator. The 920 build used the thinner Switech radiator and I added a couple of the uber thick tfc triebwerk fans to it today. It did not change the idle temp but the under load temp dropped 10 degrees!!! Yaaay
4) In fact I highly recommend these tfc triebwerk fans for any radiator application. Yes they are hugely thick at 55 mm but that is really what it takes to force (or pull in this case) air thru the thicker radiators. These particular fans are low speed & therfore very quiet. Push-pull wouldn't really be necessary with these fans, but I already had the pushers ... so ... I kept 'em. I'll suggest a google search of "tfc triebwerk review" which will provide a number of reviews and youtubes.
5) fan filters ... doh! and duh! I just put these filters on every fan I own. I decided to try these versus foam. No solid reason as to why other than I wanted to try something between the foam and screen wire. I have pets
& lets just say that animal fur becomes infused in all things involved in cooling a PC.These particular filters come with 4 M4 X some unknown length screws ... perfect for the Black Ice but too big for the Switech radiator (in the 920 build). These tidbits would be great to be mentioned as I don't find them on any seller's web sites.
Perhaps this could be another unique to techgage review series? That would be evaluating fan filters. As an 'ngineer I think about all kinds of impossible tests ... air flow restriction being at the top of the list. Effectiveness being next, but I cannot think of a simple test for that ... maybe later.
There might be a few more thoughts, but this should be good for those thinking about a WC build.
Last edited:
Hi Psi - good read. The water cooling is a little beyond me but raises good points regarding airflow contention through a radiator and tub kinks. There is no reliable method to test filters, only real conclusion would be that the more restrictive the filter, the more dust it will catch and consequently, lower airflow - rocket science that one. As to brand reliability, foam Vs. mesh etc... difficult to test. This new system of mine has both, so I'll be checking up on them over time.
As for the IBT testing, your conclusion stacks up with what Intel stated a while ago regarding HT. They recommend disabling HT as this can cause the OS to interfere with the thread handling. IBT is tuned to utilize real cores. I can't link to the statement since i can't find it, buried in a forum somewhere in the 'net.
If the math heavy nature of the workload is highly consecutive (as maths usually is), the extra threads are unlikely to help.
Testing the IMC (memory controller) as a bottleneck might be done by getting a socket 1156 i7 with 8 or 16GB of ram and running the same work on both machines then compare times. This will test the dual Vs. tri memory controller configuration with your current system. Each memory channel has it's own IMC, so 3 on the 1366 and 2 on the 1156. If the lack of the extra controller significantly slows things down, then there's your bottleneck.
Getting hold of some Xeon workstations would be a nice expensive venture for testing, but i'm not sure they would help either.
Sandy Bridge with it's better pre-fetch capabilities would be another test, are things bandwidth related Vs. efficient guesstimation, etc. If the software is regularly updated and makes use of floating point calculations, the introduction of the AVX instruction set with SB could be another avenue of performance gains.
At this point I'm just talking out of my rear, since most of this math related stuff is beyond me (a lot of things are these days...). But without testing, it's hard to confirm anything. All i could suggest is try and send one of us with a test workload or we could talk to the Dev's and do some testing with our hardware - though they probably do their own testing.
As for the IBT testing, your conclusion stacks up with what Intel stated a while ago regarding HT. They recommend disabling HT as this can cause the OS to interfere with the thread handling. IBT is tuned to utilize real cores. I can't link to the statement since i can't find it, buried in a forum somewhere in the 'net.
If the math heavy nature of the workload is highly consecutive (as maths usually is), the extra threads are unlikely to help.
Testing the IMC (memory controller) as a bottleneck might be done by getting a socket 1156 i7 with 8 or 16GB of ram and running the same work on both machines then compare times. This will test the dual Vs. tri memory controller configuration with your current system. Each memory channel has it's own IMC, so 3 on the 1366 and 2 on the 1156. If the lack of the extra controller significantly slows things down, then there's your bottleneck.
Getting hold of some Xeon workstations would be a nice expensive venture for testing, but i'm not sure they would help either.
Sandy Bridge with it's better pre-fetch capabilities would be another test, are things bandwidth related Vs. efficient guesstimation, etc. If the software is regularly updated and makes use of floating point calculations, the introduction of the AVX instruction set with SB could be another avenue of performance gains.
At this point I'm just talking out of my rear, since most of this math related stuff is beyond me (a lot of things are these days...). But without testing, it's hard to confirm anything. All i could suggest is try and send one of us with a test workload or we could talk to the Dev's and do some testing with our hardware - though they probably do their own testing.
Psi*
Tech Monkey
Add to the list below;
6) find a supplier near you or buy online an assortment of metric screws that will fit your radiator of choice. Some fans come with screws as do some radiators. The tfc triebwerk fans fans came with 2 different screws M4 & M3? or M2? ... not sure about the size of the smaller ones. The M4s worked fine with the Black Ice radiator & the smallers ones worked with the Swiftech radiator. But playing with mounting options can be limited by the length of the screws availalbe. A huge assortment is not necessary, and if the online docs for your chosen radiator are sufficient ... as in what size the tapped screw holes are ... then you only need to find that particular size.
6) find a supplier near you or buy online an assortment of metric screws that will fit your radiator of choice. Some fans come with screws as do some radiators. The tfc triebwerk fans fans came with 2 different screws M4 & M3? or M2? ... not sure about the size of the smaller ones. The M4s worked fine with the Black Ice radiator & the smallers ones worked with the Swiftech radiator. But playing with mounting options can be limited by the length of the screws availalbe. A huge assortment is not necessary, and if the online docs for your chosen radiator are sufficient ... as in what size the tapped screw holes are ... then you only need to find that particular size.