New 3820 +$2000 build

[Dan_D]

I played around with the strap settings on my 3930K for shits and giggles and could never get it rock stable. probably needs a boost in a different voltage than I am used to playing with when it comes to multiplier overclocking.

I've seen a lot of X79 boards that just didn't play nice with strap adjustments. There were a few that did, but more that didn't. ASRock actually seemed to be the most consistent about allowing me to successfully adjust that. The Rampage IV Extreme wouldn't do it until after a BIOS update. Now I can do it on my 3930K just fine. At least to 125MHz that is.

 

[Zarathustra[H]]

I've seen a lot of X79 boards that just didn't play nice with strap adjustments. There were a few that did, but more that didn't. ASRock actually seemed to be the most consistent about allowing me to successfully adjust that. The Rampage IV Extreme wouldn't do it until after a BIOS update. Now I can do it on my 3930K just fine. At least to 125MHz that is.

Ahh, Cool.

I tried this on my Asus P9X79 WS, because I wanted to test and see if I could use my Corsair Dominator GT DDR3-2000 ram at XMP settings at 2000 rather than 1866, but it was never stable for me, even with the latest BIOS.

It would boot, it just wouldn't pass stress testing.

I didn't get too upset over it though, as going from 1866mhz to 2000mhz on the RAM would have a marginal impact at best.

 

[Dan_D]

Ahh, Cool.

I tried this on my Asus P9X79 WS, because I wanted to test and see if I could use my Corsair Dominator GT DDR3-2000 ram at XMP settings at 2000 rather than 1833, but it was never stable for me, even with the latest BIOS.

It would boot, it just wouldn't pass stress testing.

I didn't get too upset over it though, as going from 1833mhz to 2000mhz on the RAM would have a marginal impact at best.

You won't get much of a tangible benefit going past DDR3 1600MHz. I used to run the latest and the fastest RAM, but half the time I had trouble getting motherboards to play nice with it. And really, I couldn't usually tell any kind of performance difference between DDR2 800MHz and 1000MHz modules etc. X58 still has more memory bandwidth than Z77 will ever have and Z77 is the faster platform overall. This tells us that memory bandwidth isn't a serious hurdle to performance at present.

 

[Zarathustra[H]]

You won't get much of a tangible benefit going past DDR3 1600MHz. I used to run the latest and the fastest RAM, but half the time I had trouble getting motherboards to play nice with it. And really, I couldn't usually tell any kind of performance difference between DDR2 800MHz and 1000MHz modules etc. X58 still has more memory bandwidth than Z77 will ever have and Z77 is the faster platform overall. This tells us that memory bandwidth isn't a serious hurdle to performance at present.

Agreed.

I don't think the extra memory bandwidth of triple or even quad channel RAM does much at all for most typical loads, especially at 4 cores or below.

I have noticed pretty poor scaling with multiple cores on the 3930k though (my own testing below)

I was expecting it to scale pretty well up to 6 cores, and then fall off dramatically once we got into the HT territory of core 7 through 12,

These results surprised me though. Scaling drops off already at 4 cores, and then continues to drop off linearly from cores 5 through 12 even though you are adding "fake" cores.

This suggests to me something else is limiting performance. It could just be Amdahl's law, but I'm betting on other subsystems limiting performance as well.

Maybe the extra memory bandwidth in X58 and X79 is intended to address situations where all 12 logical cores are loaded, but you don't really notice a difference in typical use?

 

[Dan_D]

With few exceptions, it's hard to leverage more than 4 cores right now because that's the sweet spot for software developers. That's why we don't have 8 core processors from Intel in the regular desktop line.

 

[Zarathustra[H]]

With few exceptions, it's hard to leverage more than 4 cores right now because that's the sweet spot for software developers. That's why we don't have 8 core processors from Intel in the regular desktop line.

Agreed, but Cinebench is supposed to be the best free test of multithreaded performance.

Utterly irrelevant to real world loads, I know, but I tested out of curiosity.

 

[Zarathustra[H]]

I used to run the latest and the fastest RAM, but half the time I had trouble getting motherboards to play nice with it. And really, I couldn't usually tell any kind of performance difference between DDR2 800MHz and 1000MHz modules etc.

Yeah, In retrospect buying the Corsair Dominator GT DDR3-2000 modules was a waste.

I got them because in my last build (i7-920) i had a couple of less capable Gskill sticks that held back my overclock, and I wanted to make sure that wasn't going to happen again.

 

[Dan_D]

Yeah, In retrospect buying the Corsair Dominator GT DDR3-2000 modules was a waste.

I got them because in my last build (i7-920) i had a couple of less capable Gskill sticks that held back my overclock, and I wanted to make sure that wasn't going to happen again.

X58 still allows overclocking through the bus clock. X79 / Z77 don't. At least not realistically. So memory dividers are always employed to get your RAM clocks where you want them. In the X58 days, RAM could hold your OC back. Now, not really. I ran DDR3 2000MHz modules on my X58 setup for this reason even though my OC had me running 1866MHz most of the time.

 

[Zarathustra[H]]

X58 still allows overclocking through the bus clock. X79 / Z77 don't. At least not realistically. So memory dividers are always employed to get your RAM clocks where you want them. In the X58 days, RAM could hold your OC back. Now, not really. I ran DDR3 2000MHz modules on my X58 setup for this reason even though my OC had me running 1866MHz most of the time.

Exactly. With x58 (and any system) you have to buy your RAM before you know exactly what the sweet spot for your CPU and motherboard is going to be, so you err on the side of safety and get the fastest RAM you can.

I already had 8GB of this RAM in my Phenom II which I had bought for that reason.

I went i7-920 -> Phenom II X6 -> i7-3930K.

The Phenom II was 990FX Bulldozer-ready build with a temporary Phenom II, abandoned when Bulldozer turned out to be the disappointment of the decade. I wanted to support AMD, but I'm not going to do it at the cost of having a really disappointing CPU.

Anyway, I carried the RAM forward to my X79 build, but I needed two more sticks for quad channel, so I bought two more of the same kind to make sure they played nice together.

Last week I picked up 4 more sticks of the same kind, as I had a need to create a 16GB RamDisk. I didn't feel I needed the Corsair Dominator GT's, but in the past I've learned my lesson about different memory modules not playing nice with each other, so I got more of the same.

 

[Dan_D]

Exactly. With x58 (and any system) you have to buy your RAM before you know exactly what the sweet spot for your CPU and motherboard is going to be, so you err on the side of safety and get the fastest RAM you can.

I already had 8GB of this RAM in my Phenom II which I had bought for that reason.

I went i7-920 -> Phenom II X6 -> i7-3930K.

The Phenom II was 990FX Bulldozer-ready build with a temporary Phenom II, abandoned when Bulldozer turned out to be the disappointment of the decade. I wanted to support AMD, but I'm not going to do it at the cost of having a really disappointing CPU.

Anyway, I carried the RAM forward to my X79 build, but I needed two more sticks for quad channel, so I bought two more of the same kind to make sure they played nice together.

Last week I picked up 4 more sticks of the same kind, as I had a need to create a 16GB RamDisk. I didn't feel I needed the Corsair Dominator GT's, but in the past I've learned my lesson about different memory modules not playing nice with each other, so I got more of the same.

What voltage is that RAM? My Dominators were 1.65v and I'd never dream of using that with Sandy Bridge or Ivy Bridge for extended periods of time. The Corsair Vengeance stuff I have is 1.5v at 1600MHz.

 

[Zarathustra[H]]

What voltage is that RAM? My Dominators were 1.65v and I'd never dream of using that with Sandy Bridge or Ivy Bridge for extended periods of time. The Corsair Vengeance stuff I have is 1.5v at 1600MHz.

It's 1.65v Never had an issue with it, and from what I read, 1.65v is within safe specifications for SB, SB-E and IB, even though for some reason many claim otherwise.

According to Intel's own documentation Vccdmax is 1.85v, so 1.65v is well within the spec.

I feel like all the talk of Sandy bridge and Ivy bridge requiring 1.5v RAM is a little bit or urban legend.

 

[Dan_D]

It's 1.65v Never had an issue with it, and from what I read, 1.65v is within safe specifications for SB, SB-E and IB, even though for some reason many claim otherwise.

According to Intel's own documentation Vccdmax is 1.85v, so 1.65v is well within the spec.

I feel like all the talk of Sandy bridge and Ivy bridge requiring 1.5v RAM is a little bit or urban legend.

I remembered seeing some specifications from Intel stating that going over 1.6v for Sandy Bridge was bad. I don't know I guess I remembered the numbers wrong.

 

[Zarathustra[H]]

I remembered seeing some specifications from Intel stating that going over 1.6v for Sandy Bridge was bad. I don't know I guess I remembered the numbers wrong.

Well, it's important to read the fine print.

If we translate the lawyerish engineering speak they more or less state in the above that 1.85v is the max safe voltage for short stints, but long term use may degrade the CPU memory controller.

I don't recall Intel releasing any official max specs for regular Sandy bridge, but the official recommendation is 1.5v, and I think that's why everyone is treating it like religion.

There are plenty of people running 1.65v RAM on Sandy bridge for extended periods of time though, so I am comfortable with it. I'm not sure I would push any higher though. That might be playing with fire.

 

[Dan_D]

Well, it's important to read the fine print.

If we translate the lawyerish engineering speak they more or less state in the above that 1.85v is the max safe voltage for short stints, but long term use may degrade the CPU memory controller.

I don't recall Intel releasing any official max specs for regular Sandy bridge, but the official recommendation is 1.5v, and I think that's why everyone is treating it like religion.

There are plenty of people running 1.65v RAM on Sandy bridge for extended periods of time though, so I am comfortable with it. I'm not sure I would push any higher though. That might be playing with fire.

Well I hadn't really held to it like it was a religion. Something I vaguely remember. People also said not to run 1.7v DDR 3 RAM with X58 and I did it for some time before 1.65v and lower enthusiast modules came out. I figured that it wasn't a hard line as I've also run my Corsair Dominators on Sandy Bridge before.

If you've got some really good 1.65v RAM laying around that you can reuse, then by all means. If you are buying new RAM, there is no reason to buy anything over 1.5v unless you are a benchmarking whore and want RAM that runs at DDR3 2133MHz+ speeds in which case everything I've seen (though I haven't looked much) that's at DDR3 2000+ speeds is all over 1.5v.

 

[cemster]

@LigTasm Gawd Thank you for the info.

@Dan_D

The Cire i7 3960X and Core i7 3930K are multiplier unlocked. It's only in multi-monitor and multi-GPU systems on the ultra high end of the scale where clock speeds make that much of a difference.

Thank you for the explanation. I appreciate it.
I do need a multi-monitor system actually. At least three of them.

I'd go for a Corsair H80 or H100 for cooling and never worry about heat on a 3820. Ever.

As far as 3960x goes do you think H100 is something I should prefer over a custom loop?

I do need a super silent rig. And I want to indulge myself with bunch of water cooling stuff.
I've been studying overclock.net for options and that place may not the best place to get ideas on a quiet rig! Like there is a build log, whose primary goal is Quiet! Ad he is planning on between 4 and 6 large radiators, with about 40 fans spinning at very low RPM!!!

There are some fanless PSUs for example, but they are low on power so I was wondering if I could use two of them to achieve enough juice to power everything... Or if I could use water loops for everything then place the Radiators and fans away from the case!
Well I guess I am not even ready to shape my questions properly yet!

Back to my original concern about ram being 8x channel... I've heard that LGA 2011 is not very forgiving with mixed ram sets... So I will stick with 8x set. But should I settle with 1600MHz or go for extra umph?

 

[cyclone3d]

@LigTasm Gawd Thank you for the info.
Back to my original concern about ram being 8x channel... I've heard that LGA 2011 is not very forgiving with mixed ram sets... So I will stick with 8x set. But should I settle with 1600MHz or go for extra umph?

Personally, I would go with the fastest you can get... Not necessarily the fastest stock rated.

Whatever RAM you do get, try to look for the lowest latency rating at the given speed. Even if you have to run at a reducled clock rate when filling all 8 slots, it should also mean that you will be able to tighten the timings up quite a bit.

The tighter the timings, the smoother the whole system is going to feel.

Take a look at this link to give you a good idea of how timings and speed affects the LGA 2011 platform.

http://www.techpowerup.com/reviews/Samsung/MV-3V4G3/7.html

Some of the other pages in this review also show some interesting things.

 

[Dan_D]

@LigTasm Gawd Thank you for the info.

@Dan_D

Thank you for the explanation. I appreciate it.
I do need a multi-monitor system actually. At least three of them.

As far as 3960x goes do you think H100 is something I should prefer over a custom loop?

I do need a super silent rig. And I want to indulge myself with bunch of water cooling stuff.
I've been studying overclock.net for options and that place may not the best place to get ideas on a quiet rig! Like there is a build log, whose primary goal is Quiet! Ad he is planning on between 4 and 6 large radiators, with about 40 fans spinning at very low RPM!!!

There are some fanless PSUs for example, but they are low on power so I was wondering if I could use two of them to achieve enough juice to power everything... Or if I could use water loops for everything then place the Radiators and fans away from the case!
Well I guess I am not even ready to shape my questions properly yet!

Back to my original concern about ram being 8x channel... I've heard that LGA 2011 is not very forgiving with mixed ram sets... So I will stick with 8x set. But should I settle with 1600MHz or go for extra umph?

My Koolance Exos 2.5 is far more powerful as a cooling solution than the H100 ever will be. The same can be said for any decent water cooling equipment. The Corsair H100 is really just an alternative to traditional heat pipe based air coolers as it goes a single step further in cooling. The real attractive thing about these is their ease of installation, their look, and their simplicity. Their performance is better than that of most air coolers, but it's not in the same league as more traditional water cooling loops are.

Unfortunately, power and silence do not generally go together. The most powerful cooling solutions will be louder than ones that are silent or near silent. The same is true of video cards and power supplies, or any fan. That being said a balance can always be struck. I keep my Koolance Exos on a speed setting of 6. It goes to 10 and gets a hell of a lot louder for it, but the kooling power doesn't increase dramatically for setting it that high. I get 3 or 4c off my temps at best. If I were going for a super overclock all the time this might make some sense, but 100MHz or so that I'd get from this extra headroom isn't worth experiencing the jet turbine like noise all the time. 6 can definitely be heard, but it sounds more like a quiet push of air rather than something more abrasive to my ears.

An H100 can do the job of a high end air cooler while putting less stress on your motherboard and with less noise as water beats heatpipes almost every time. Crank the fans of an H100 all the way up or replace them with something stronger and it will out cool all the air coolers out there. Though it won't really be any better than high performance air coolers noise wise at that point.

 

[Dan_D]

Personally, I would go with the fastest you can get... Not necessarily the fastest stock rated.

Whatever RAM you do get, try to look for the lowest latency rating at the given speed. Even if you have to run at a reducled clock rate when filling all 8 slots, it should also mean that you will be able to tighten the timings up quite a bit.

The tighter the timings, the smoother the whole system is going to feel.

Take a look at this link to give you a good idea of how timings and speed affects the LGA 2011 platform.

http://www.techpowerup.com/reviews/Samsung/MV-3V4G3/7.html

Some of the other pages in this review also show some interesting things.

Maybe I'm missing something, but that shows the opposite of what your stating. The timings are being scaled with the clock speeds in those tests and the performance is going up. That doesn't lend a lot of credence to the idea that latencies are more important than clocks. The only test where this differs is the latency test which shows a slight performance advantage over tighter timings rather than clock speed, but that's a synthetic test which is designed to test latencies. So I'm not surprised.

You aren't going to be able to tell the difference between 9,9,9,24,1T vs. 10,10,10,27,1T outside of a benchmark. Sorry but they are going to "feel" the same.

 

[Zarathustra[H]]

Maybe I'm missing something, but that shows the opposite of what your stating. The timings are being scaled with the clock speeds in those tests and the performance is going up. That doesn't lend a lot of credence to the idea that latencies are more important than clocks. The only test where this differs is the latency test which shows a slight performance advantage over tighter timings rather than clock speed, but that's a synthetic test which is designed to test latencies. So I'm not surprised.

You aren't going to be able to tell the difference between 9,9,9,24,1T vs. 10,10,10,27,1T outside of a benchmark. Sorry but they are going to "feel" the same.

Agreed.

Anand's article is maybe even more clear on this subject.

Bandwidth seems to matter more than timings in most tests, but in the real world tests the fastest (DDR3-2133 9-11-9-27) is pretty close to the slowest (DDR3-1333 9-9-9-18), with only a few tests showing a practical difference between the two.

 

[Dan_D]

Agreed.

Anand's article is maybe even more clear on this subject.

Bandwidth seems to matter more than timings in most tests, but in the real world tests the fastest (DDR3-2133 9-11-9-27) is pretty close to the slowest (DDR3-1333 9-9-9-18), with only a few tests showing a practical difference between the two.

This is why I went with DDR3 1600MHz memory in this build. There just wasn't a compelling reason to go with faster RAM in practice. I'm not a benchmark whore, so yeah. Not worth it.

 

[cyclone3d]

Maybe I'm missing something, but that shows the opposite of what your stating. The timings are being scaled with the clock speeds in those tests and the performance is going up. That doesn't lend a lot of credence to the idea that latencies are more important than clocks. The only test where this differs is the latency test which shows a slight performance advantage over tighter timings rather than clock speed, but that's a synthetic test which is designed to test latencies. So I'm not surprised.

You aren't going to be able to tell the difference between 9,9,9,24,1T vs. 10,10,10,27,1T outside of a benchmark. Sorry but they are going to "feel" the same.

I think I maybe didn't word it the best way I could have.

For any given speed, the tighter the timings, the smoother it is going to feel because latency is going to be lower. So even if when all 8 slots are filled up, and having to run the RAM at a lower speed, the tighter timings vs lower speed rated RAM should help make up for it a bit.

The main timings aren't the only thing to think about either. Tightening up the sub-timings can also give around 1GB/s or more in throughput increase as well as affect the overall latency.

Once my x79 parts get here, I will be able to get a better handle on how the new setups work since I am still on the x58 platform right now.

 

[Zarathustra[H]]

I think I maybe didn't word it the best way I could have.

For any given speed, the tighter the timings, the smoother it is going to feel because latency is going to be lower. So even if when all 8 slots are filled up, and having to run the RAM at a lower speed, the tighter timings vs lower speed rated RAM should help make up for it a bit.

The main timings aren't the only thing to think about either. Tightening up the sub-timings can also give around 1GB/s or more in throughput increase as well as affect the overall latency.

Once my x79 parts get here, I will be able to get a better handle on how the new setups work since I am still on the x58 platform right now.

I'm sorry, but if you are feeling any difference at all when going from 55.2ns (slowest in your linked test) to 43.9ns (fastest in your linked test) it is likely all in your head (placebo effect).

We are talking nanoseconds here. A difference of 11.3 billionths of a second is not humanly detectable.

That's 0.0000113ms or 0.0000000113s.

Its very possible you are detecting a difference, but that is only because you think there is one, and your brain is fooling you.

Never trust your (or anyone else's) brain. The human brain simply isn't that reliable, and will make you experience something, just because you think it is happening, whether it is real or not. This is why we do double blinded studies in clinical trials for medications. You can't trust the brains of either the scientists, clinicians or patients. They are all guaranteed to experience bias that makes them honestly think they are seeing or feeling what they expect to see or feel.

 

[kirbster]

I tried straps on the hex cores as well and they don't seem to like them very much. The 3820 loves 125 though, I've used two and both made it over 5ghz easily.

I got mine to 5.2Ghz on a vanilla P9X79.... 125 strap is a no brainer.

 

[Dan_D]

I think I maybe didn't word it the best way I could have.

For any given speed, the tighter the timings, the smoother it is going to feel because latency is going to be lower. So even if when all 8 slots are filled up, and having to run the RAM at a lower speed, the tighter timings vs lower speed rated RAM should help make up for it a bit.

The main timings aren't the only thing to think about either. Tightening up the sub-timings can also give around 1GB/s or more in throughput increase as well as affect the overall latency.

Once my x79 parts get here, I will be able to get a better handle on how the new setups work since I am still on the x58 platform right now.

I'm sorry, but if you are feeling any difference at all when going from 55.2ns (slowest in your linked test) to 43.9ns (fastest in your linked test) it is likely all in your head (placebo effect).

We are talking nanoseconds here. A difference of 11.3 billionths of a second is not humanly detectable.

That's 0.0000113ms or 0.0000000113s.

Its very possible you are detecting a difference, but that is only because you think there is one, and your brain is fooling you.

Never trust your (or anyone else's) brain. The human brain simply isn't that reliable, and will make you experience something, just because you think it is happening, whether it is real or not. This is why we do double blinded studies in clinical trials for medications. You can't trust the brains of either the scientists, clinicians or patients. They are all guaranteed to experience bias that makes them honestly think they are seeing or feeling what they expect to see or feel.

This. And if you aren't running X79 now, you really don't know what you are talking about. I've said it before and I'll say it again, the Intel platform just isn't that sensitive to timings. And clock frequencies and therefore bandwidth aren't as important as you think they are. Otherwise there would be a huge difference in performance between LGA1155 / Z77 Express and LGA2011 / X79 Express. In games and in most applications, there isn't. X58 had more bandwidth than Z77 does and it's slower, even with Gulftown on it.

 

[needmorecarnitine]

The only feeling that I would detect after buying the fastest ram would be the feeling of getting fucked for paying too much for RAM

 

[Dan_D]

The only feeling that I would detect after buying the fastest ram would be the feeling of getting fucked for paying too much for RAM

I'm all for going fast but at some point you reach the point of diminishing returns. When it comes to RAM, currently we can reach that point very quickly.

 

[cemster]

So what I understand from these tests is that neither the bandwidth over 1600MHz, nor the lower timings are matter as far as X79 goes... Correct?

OH! I see we moved to a new page...

The Intel platform just isn't that sensitive to timings. And clock frequencies and therefore bandwidth aren't as important as you think they are.

And I did understand correctly... )
Thanks to you guys...

So consequently "CORSAIR Vengeance 64GB (8 x 8GB) 240-Pin DDR3 SDRAM DDR3 1600 " is good enough for the Mobo/CPU I'm in...

 

[Dan_D]

So what I understand from these tests is that neither the bandwidth over 1600MHz, nor the lower timings are matter as far as X79 goes...

Yep. Keep in mind that even at stock clocks the X79 platform has more bandwidth than Z77 and X58 do. X79 based systems aren't necessarily faster in games and in many cases are even slower due to lower clocks. Sandy Bridge and Ivy Bridge tend to clock a little better than Sandy Bridge-E does. Probably because you've got 6 cores instead of 4. And Gulftown at 4.4GHz on X58 isn't faster than Z77 despite having about 33% more memory bandwidth. So again bandwidth isn't doing anything in games. In fact the applications that benefit most from the bandwidth that X79 offers will be content creation, SQL, database applications, etc. In other words, stuff for the server and workstation market. Not games.

In fact X79 and Sandy Bridge-E largely target those markets by design. There is some overlap with the high end / enthusiast desktop market, but those customers are rarely better served by X79 over Z77. In some cases they'd be better off with Z77.