5960x with H110 cooler on a MSI Krait motherboard - anyone have overclock parameters

[proggy123]

Hi, as the post says, i'm running a 5960x with a H110 cooler on a MSI Krait SLI motherboard. Ram is 32 gb of Corsair Dominator 3000 - i've kept that at 2133 for now.

I've been playing around with overclock settings and i think i've managed to get it to 3.9 stable(ish...). But people seem to be getting to 4.3-4.5 without any trouble.

I was wondering if anyone could share information or any sites that might have bios parameters i could try to achieve a higher stable overclock. 4.1 or 4.2 would be very nice if i could achieve it and still have a machine that i only need to reboot once a month or less (i'm running linux).

Any advice would be appreciated.

Also, anyone running very large amounts of RAM ? I'm looking at getting 128 GB for this machine but i'm worried that without ECC it'll be practically unusable because of bit errors.

Interestingly, for the first 2-3 months since i built this box, i would get a ton of memory errors, but now i don't have any noticeable amount - i wonder if RAM needs burn-in or something.

 

[Burpo]

Have you tried their OC software, MSI Command Center Lite? It's free from MSI and may help you dial in your settings.

"CC Lite is now on v2, suitable for X99 processors. Here adjusting the text and clicking Apply is all that is needed to enable an overclock. Users can also adjust the overclock with a selected hotkey, which is especially useful for benchmarks that have variable CPU load sections."

http://www.anandtech.com/show/8557/...ck-x99-ws-msi-x99s-sli-plus-intel-haswell-e/9

 

[alyarb]

Are you trying to have a conversation about "4.3-4.5" without asking about voltage?

 

[proggy123]

Thanks Burpo - no i hadn't seen that. I'll definitely check it out.

Alyarb, no, i didn't have any ulterior motive of deliberately not asking about voltage or any other factors that might matter. Sorry if my post gave that impression. I'm not prejudiced against voltage and am happy to talk about it along with anything else you feel would help my cause.

 

[Techhog]

Thanks Burpo - no i hadn't seen that. I'll definitely check it out.

Alyarb, no, i didn't have any ulterior motive of deliberately not asking about voltage or any other factors that might matter. Sorry if my post gave that impression. I'm not prejudiced against voltage and am happy to talk about it along with anything else you feel would help my cause.

... What?


Uh... Just tell us what you set your voltage to. Also, I recommend against OCing a CPU with motherboard utilities. Also, if you never touched your voltage, we've found your problem.

Also, if you actually need 128GB of RAM, you're probably doing serious work and shouldn't overclock in the first place. What do you feel you need all of that RAM for?

 

[proggy123]

I was following this guide - and here's the page i got voltage advice from.

http://www.tweaktown.com/guides/675...-spec-qfra-cpu-overclocking-guide/index4.html -
Right now, my Vin is set to 2.0 with Vcore at 1.28 i believe. I can post the rest of my settings when i get home today. But that guide is for a different motherboard and mine doesn't have the exact same settings hence my questions about whether anyone had any data on the Krait SLI board.

Boring details on workload below, sorry if it's tmi.
Re. the 128 gb ram vs. overclock question - that's a good one. The workload is mostly simulations - various statistical and/or exact models. The boring details is are that the code is c++, numa-aware and reasonably optimized for speed in terms of cache-awareness and using vectorization intrinsics. I reserve one core for the kernel and the other 7 are isolated and my code runs at realtime priority on them. Turbo is off, I run with linux cpu settings set to "performance" across all cores.

For my workload it come down to how much throughput i can get per core. Going from 3 ghz to 4ghz across all cores would almost give me a 33% speedup for my workload, so i'm interested in overclocking.

Since most of the models are statistical, i do multiple runs anyway and i have cross-checks to drop outliers and see if results look egregiously out of whack. So i'm not super sensitive to one-off bit errors unless they persist across time or there are just so many that i get errors on every single run - which is why i've gone the non-ECC non-Xeon route for now (Also price...)

I'm considering 128 gb is just because it's the max ram i could get for this platform. There's a lot of overlapping computation in what i do, so caching intermediate results in ram is a huge speed win - even over ssd. Also, being able to visualize large datasets very fast because they're completely loaded in memory really helps my research process. If i could get even more ram, i would.

A lot of this stuff can be done distributed using something like Spark and eventually to run on really large datasets I end up doing just that, but frankly, running on 1 machine that's maxed out and can handle "reasonably large" quantities of data, lets me iterate on research much faster.

So that's why i've been running the 5960x with 32gb ram - it's doing very well which is why i'm considering maxing out the ram and seeing if i can get a higher stable overclock.

 

[Headfoot]

I'd suggest turning your memory XMP profile on because it may require you to run at a different bus speed, which will essentially reset your current overclock as you will need to find a new multiplier that works with the higher bus speed. Besides, what may be stable with 2133 RAM may not be stable with 3000 RAM so its best to find your real settings here. You can always turn it down to 2133 when you get instability to see if its due to that speed setting

I got my 5820k to 4.4 on ~1.25-1.28 (depending on load line calibration under load), w/ Bus at 125 instead of 100. Your chip may require more voltage or less for the same speed

 

[Madpacket]

I was following this guide - and here's the page i got voltage advice from.

http://www.tweaktown.com/guides/675...-spec-qfra-cpu-overclocking-guide/index4.html -
Right now, my Vin is set to 2.0 with Vcore at 1.28 i believe. I can post the rest of my settings when i get home today. But that guide is for a different motherboard and mine doesn't have the exact same settings hence my questions about whether anyone had any data on the Krait SLI board.

Boring details on workload below, sorry if it's tmi.
Re. the 128 gb ram vs. overclock question - that's a good one. The workload is mostly simulations - various statistical and/or exact models. The boring details is are that the code is c++, numa-aware and reasonably optimized for speed in terms of cache-awareness and using vectorization intrinsics. I reserve one core for the kernel and the other 7 are isolated and my code runs at realtime priority on them. Turbo is off, I run with linux cpu settings set to "performance" across all cores.

For my workload it come down to how much throughput i can get per core. Going from 3 ghz to 4ghz across all cores would almost give me a 33% speedup for my workload, so i'm interested in overclocking.

Since most of the models are statistical, i do multiple runs anyway and i have cross-checks to drop outliers and see if results look egregiously out of whack. So i'm not super sensitive to one-off bit errors unless they persist across time or there are just so many that i get errors on every single run - which is why i've gone the non-ECC non-Xeon route for now (Also price...)

I'm considering 128 gb is just because it's the max ram i could get for this platform. There's a lot of overlapping computation in what i do, so caching intermediate results in ram is a huge speed win - even over ssd. Also, being able to visualize large datasets very fast because they're completely loaded in memory really helps my research process. If i could get even more ram, i would.

A lot of this stuff can be done distributed using something like Spark and eventually to run on really large datasets I end up doing just that, but frankly, running on 1 machine that's maxed out and can handle "reasonably large" quantities of data, lets me iterate on research much faster.

So that's why i've been running the 5960x with 32gb ram - it's doing very well which is why i'm considering maxing out the ram and seeing if i can get a higher stable overclock.

AFAIK max RAM allowed for 5960X is 64GB.

 

[myocardia]

AFAIK max RAM allowed for 5960X is 64GB.

That would be correct, at least according to Intel. While the "platform" does indeed support 128GB of ECC RAM, the i7's support neither of the two. If you want either, you have to buy a Xeon.

 

[Techhog]

That would be correct, at least according to Intel. While the "platform" does indeed support 128GB of ECC RAM, the i7's support neither of the two. If you want either, you have to buy a Xeon.

128GB should still boot, though.

 

[Techhog]

I was following this guide - and here's the page i got voltage advice from.

http://www.tweaktown.com/guides/675...-spec-qfra-cpu-overclocking-guide/index4.html -
Right now, my Vin is set to 2.0 with Vcore at 1.28 i believe. I can post the rest of my settings when i get home today. But that guide is for a different motherboard and mine doesn't have the exact same settings hence my questions about whether anyone had any data on the Krait SLI board.

Boring details on workload below, sorry if it's tmi.
Re. the 128 gb ram vs. overclock question - that's a good one. The workload is mostly simulations - various statistical and/or exact models. The boring details is are that the code is c++, numa-aware and reasonably optimized for speed in terms of cache-awareness and using vectorization intrinsics. I reserve one core for the kernel and the other 7 are isolated and my code runs at realtime priority on them. Turbo is off, I run with linux cpu settings set to "performance" across all cores.

For my workload it come down to how much throughput i can get per core. Going from 3 ghz to 4ghz across all cores would almost give me a 33% speedup for my workload, so i'm interested in overclocking.

Since most of the models are statistical, i do multiple runs anyway and i have cross-checks to drop outliers and see if results look egregiously out of whack. So i'm not super sensitive to one-off bit errors unless they persist across time or there are just so many that i get errors on every single run - which is why i've gone the non-ECC non-Xeon route for now (Also price...)

I'm considering 128 gb is just because it's the max ram i could get for this platform. There's a lot of overlapping computation in what i do, so caching intermediate results in ram is a huge speed win - even over ssd. Also, being able to visualize large datasets very fast because they're completely loaded in memory really helps my research process. If i could get even more ram, i would.

A lot of this stuff can be done distributed using something like Spark and eventually to run on really large datasets I end up doing just that, but frankly, running on 1 machine that's maxed out and can handle "reasonably large" quantities of data, lets me iterate on research much faster.

So that's why i've been running the 5960x with 32gb ram - it's doing very well which is why i'm considering maxing out the ram and seeing if i can get a higher stable overclock.

Is that speed worth the possible instability and data corruption, though? If your workload ends up crashing and taking data with it, you've lost time. On top of that, scaling isn't linear. (Also, the max turbo on 8 cores with a 5960X is 3.3GHz, so you'd actually see a 21% boost at best.) Personally, it sounds to me like you'd have been better off buying a Xeon with more cores and ECC support.

 

[Madpacket]

128GB should still boot, though.

I have a 5930K running at 4.2Ghz (this chip is a dud overclocking wise) + MSI X99 Raider (very similar to the Krait) with 4X16GB DDR4 G.Skill 3000Mhz RAM. I still have 4 slots free and would love to test out 128GB. I purposely stopped at 64GB as I figured even if it boots only 64GB will be addressable due to an artificial limitation imposed by Intel but but I would love to be wrong here. I'm not spending another 600+ to test out this theory at least until Broadwell Xeon's are available

Is that speed worth the possible instability and data corruption, though? If your workload ends up crashing and taking data with it, you've lost time. On top of that, scaling isn't linear. (Also, the max turbo on 8 cores with a 5960X is 3.3GHz, so you'd actually see a 21% boost at best.) Personally, it sounds to me like you'd have been better off buying a Xeon with more cores and ECC support.

Although I think you're right in principle ECC memory and Intel Xeon's are quite a bit more expensive for less performance so I get what the OP is trying to do here (save money!) and still have good performance.

Re overclocking, I don't see how you risk having more memory errors due to CPU overclocking. Leave the RAM speed alone and just change the multiplier and work on stabilizing the CPU. Regardless of what speed you end up at - Stable means Stable. If you have a stable CPU even when overclocked you shouldn't have more memory errors. These factors are mostly independent of each other. Statistically memory errors should stay the same unless you add more memory. ECC is seriously overrated for most workloads and a waste of money. There are other factors at play here as well. How does your simulation / program currently handle memory errors? Do you factor for these errors in your program or just rely on multiple runs and throw out the anomalies? Are you sure the anomalies are caused by memory errors? Does recent Linux kernel memory management handle memory errors more gracefully?

Is ECC really necessary for your workload? This is something that would be good to test out but I suspect it wouldn't make much of a difference.

 

[Burpo]

128Gb isn't a problem for Haswell-E..

"..Corsair for its 128GB kit. The company actually has three speeds: DDR4/2800, DDR4/2666 and DDR4/2400 at the Platinum level. The kit we used was the DDR4/2400 and is comprised of eight 16GB modules"

http://www.pcworld.com/article/2938...s-pc-with-128gb-of-cutting-edge-ddr4-ram.html

 

[Madpacket]

Also, anyone running very large amounts of RAM ? I'm looking at getting 128 GB for this machine but i'm worried that without ECC it'll be practically unusable because of bit errors.

Interestingly, for the first 2-3 months since i built this box, i would get a ton of memory errors, but now i don't have any noticeable amount - i wonder if RAM needs burn-in or something.
proggy123 is offline Report Post

Memory either works or it doesn't. It's fairly binary, it either works stably or it causes errors. There's no burn in per sae that you need to do.

Have you run the latest memtest to ensure it passes multicore testing over 24 hours? What about Prime 95 Large FFT with check-sums enabled? Does it pass at default speeds?

First you need to determine the memory you have is stable. From what you described it sounds like you may already have bad memory. These two programs will likely stress out your memory much more than your program ever will.

As for the need for ECC, it's a little complicated.

What altitude are you at? (serious question). The higher the altitude the greater the chance of cosmic rays causing bit flips in memory. Is your workstation shielded properly?

I (personally) think with modern operating systems and better voltage regulation and manufacturing the need for ECC is really overrated but I'm not running your program to do a comparison.

 

[Madpacket]

128Gb isn't a problem for Haswell-E..

"..Corsair for its 128GB kit. The company actually has three speeds: DDR4/2800, DDR4/2666 and DDR4/2400 at the Platinum level. The kit we used was the DDR4/2400 and is comprised of eight 16GB modules"

Wow thank you for that link. So if I have BIOS support or with simple BIOS update I should be able to top out my system with another 4*16GB bringing me up to 128GB....

Need to confirm with MSI but if my board supports it I will be buying more RAM pretty soon ;D

 

[Burpo]

That's an old article.. Pretty sure all the new bios (s) supports 16Gb sticks..

 

[Madpacket]

That's an old article.. Pretty sure all the new bios (s) supports 16Gb sticks..

Considering I'm already using 16GB sticks this makes sense but not sure if the BIOS is setup to read more than 64GB. Hmm..

 

[proggy123]

Thanks everyone for all the advice. @Madpacket, i think you're right - my corsair memory is probably bad which is why i'm getting bit errors. Sigh, too late to return now probably. Time to call Corsair after running memtest. But say 1% error-rate is fine with me - when i'm reading pre-cached data, i compute checksums so i know immediately if data was corrupted. During computation, each run is anywhere from 2-20 minutes, but there are check steps mingled with compute steps and compute steps are automatically retried a number of times if they fail. So far, i haven't had more than one or two times (out of tens of thousands) where a compute step has failed more than 2 times consistently. Which tells me there's errors but they are intermittent and don't affect the same memory locations all the time.

The kernel can't do much for me unfortunately. It'll map virtual pages to physical ram, but right now i restrict my computations to about 30 gb memory anyway. I get the ram from the kernel upfront at the beginning of my code and then use a custom allocator to shove data into it. I'm also running kernel 4.2 since i'm on ubuntu 15.10

@Techhog, i figured bit errors were a calculated risk given that i have code that alerts me to it. Also, re. your comments about Turbo you're right - which is why i run without it. Turbo iirc, actually speeds up specific cores and the turbo multiplier goes down the more cores you want sped up - which is why i'd like to overclock because that i would think would affect all cores.

@Burpo, yeah when i decided to buy into the 5960x, i did a bunch of reading and it seemed like the 5960x haswell-e would be able to handle 128gb. Admittedly i haven't upgraded my BIOS in a while so it's possible that will be a stumbling block.

Looks like there's 128gb kits for about 800$ now on newegg. Maybe in a couple of months after i get back from vacation the price will have fallen a bit more.

 

[Madpacket]

Nice. Even with ECC it's a good idea to perform checksums as errors from CPU cache and elsewhere can and will happen. Sounds like you've written a fairly resilient application that uses atomic principles so I don't see how non ECC would be much of an issue.

I almost pulled the trigger on another 64GB yesterday but the price has gone up quite a bit since I last purchased (due to falling CAD). I'll also be waiting for a bit to buy more RAM unless I run into a good sale.

 

[guskline]

I use Asus RealBench and Intel Overclocking utilities to test my overclocking stability.

Each 5960x is probably a bit different. I use an Asus Rampage V mb which is designed as a higher end mb to OC. I only have 16G of Corsair Vengence DDR4-3000 but opted to keep it at stock 2133 and tweek some memory settings to get better performance figures mesure by Aida64 benchmarking tool.

I found for me, the most solid OC was 4.4Ghz (100x44) at 1.33 Vcore. I have a custom water cooled sytem (EK Supremacy EVO block and 3 480 rads etc etc) so cooling is NOT a problem. To get to 4.5 stable I have to up the vcore to nearly 1.4 which is not worth the heat.

BTW, I ALWAYS buy the optional Intel OC insurance for my cpus if they qualify. For both my 5960x and my 4790k it allows a one time replacement if failure is due to OCing. I think the price was $35 for the 5960x and $25 for the 4790k.

If your chip can OC stable at 3.9Ghz without any change to the vcore, be happy. If you want to get into the 4.3 to 4.5 Ghz range I suspect you need to up the vcore. Your H110 cooler should be able to handle the increased heat BUT be aware as you keep upping the speed the heat output of the 8core/16 thread "monster" will take off.

With your setup, I would set a MAX vcore of 1.3 and set the multiplier to 4.1Ghz and then see how stable it is. Perhaps 4.2 or even 4.3 but watch the heat and keep an eye on the thermals.

In addition, as you increase the amount of ram, you increase overall stress on the chipset etc, thus my suggestion on starting with 1.3 vcore max and even 4.0 Ghz and see how stable it is.

Good luck.