Memory recommendation for 5930K system

[Carfax83]

I recently purchased a 5930K as I'm building a new system. I don't have all the parts yet though, the only thing I have left to get is the RAM..

I'm not sure whether I should go with a 16GB kit, or a 32GB kit. Usually I'd say 16GB is enough for high end gaming, but next year's GPUs will have very large frame buffers (even more than Titan X for certain models), and I'm wondering whether 16GB of RAM will be enough to buffer next generation Pascal GPUs, which will reportedly have as much as 16GB of VRAM on their top end consumer GPU (Titan XI perhaps), and as much as 32GB on their professional GPUs.

Another thing I'm wondering is what speed RAM to get. If I get 32GB, I would likely want to get at least DDR4 3000 speed, as I'll probably have this new system until Cannonlake-E arrives in 2018 or so..

If I opt for DDR4 3200 instead, then it would have to be 16GB because I don't think any manufacturer makes 32GB quad channel kits at that speed.

But from my reading, it seems as though Haswell-E has problems with high DDR4 speeds unlike Skylake, so I may not even get a chance to take advantage of the RAM speed..

 

[Carfax83]

OK I decided on the TridentZ DDR4-3200 32GB quad channel kit.

I thought they didn't make quad channel kits at that speed, but I was wrong. I doubt my 5930K will allow me to run at that RAM speed though, but I'll probably opt to get an 8 core Broadwell-E CPU later after they drop down in price and that shouldn't have any issues..

I figure I may as well go 32GB, since even high speed RAM is fairly inexpensive. Games are using more and more RAM, like Just Cause 3 which will use almost 10GB in a 32GB system. And while that's still quite a ways off from 16GB, that kind of RAM usage might cause problems in a 16GB system in which the pagefile is disabled whereas on a 32GB system it wouldn't cause a problem.

And I always disable my page file

 

[Termie]

What you'll find is that the only way to run your memory at that speed is to use a 125MHz strap, which distorts other parameters of the system. The idle CPU speed is 1500MHz instead of 1200MHz, power use is much higher, and the CPU ratios are very different for overclocking purposes (no longer even multiples of 100).

I was running 4x4GB of DDR4-3000 in my X99 system and then decided to switch to 4x8GB of DDR4-2666. That's the highest speed you can run without the odd strap, and it also runs at 1.2V instead of 1.35V. With the bandwidth available on X99, I don't think RAM speed matters nearly as much as on Skylake.

 

[Carfax83]

What you'll find is that the only way to run your memory at that speed is to use a 125MHz strap, which distorts other parameters of the system. The idle CPU speed is 1500MHz instead of 1200MHz, power use is much higher, and the CPU ratios are very different for overclocking purposes (no longer even multiples of 100).

I was running 4x4GB of DDR4-3000 in my X99 system and then decided to switch to 4x8GB of DDR4-2666. That's the highest speed you can run without the odd strap, and it also runs at 1.2V instead of 1.35V. With the bandwidth available on X99, I don't think RAM speed matters nearly as much as on Skylake.

Hey thanks for your feedback Termie. While I think you're right about DDR4 3000 requiring a 125MHz strap for stability, DDR4 3200 (which is what I ended up buying) is known to be stable on 100MHz strap.

In fact, G.Skill programs their DDR4 3200 modules to use strap 100MHz with XMP..

We'll see how it goes when I get all the parts in. I'll make sure to update this thread.

 

[Termie]

Hey thanks for your feedback Termie. While I think you're right about DDR4 3000 requiring a 125MHz strap for stability, DDR4 3200 (which is what I ended up buying) is known to be stable on 100MHz strap.

In fact, G.Skill programs their DDR4 3200 modules to use strap 100MHz with XMP..

We'll see how it goes when I get all the parts in. I'll make sure to update this thread.

Wow, that's new to me. I might try running my 3000/cas15 sticks at 3200/cas16 to see I can get that to work. Had I known that, I would have grabbed the same sticks you did (they were on sale over Black Friday).

 

[Carfax83]

Wow, that's new to me. I might try running my 3000/cas15 sticks at 3200/cas16 to see I can get that to work. Had I known that, I would have grabbed the same sticks you did (they were on sale over Black Friday).

What was the sale price on Black Friday? I've been out of the country for a bit so I'm curious.

Anyway, my source is Raja over at RoG forums. He knows his stuff so I have no reason to doubt him:

1) The most stable DRAM ratio over 2400, is 3200 on the 100 strap. If you use the 125 strap, the DDR4-2400 ratio becomes DDR4-3000 (which is stable using 125). All other DRAM ratios that fall between DDR4-2400 and DDR4-3200 on the 100 strap are tricky to get stable on most CPUs.

2) Using the 100 strap allows the use of adaptive voltage. But has the side-effects for DRAM listed above.

Source

 

[Termie]

What was the sale price on Black Friday? I've been out of the country for a bit so I'm curious.

Anyway, my source is Raja over at RoG forums. He knows his stuff so I have no reason to doubt him:



Source

They were $225. A Newegg mobile coupon would have knocked it down another 5%.

That's a pretty old thread. You might want to see this comment from G.skill on Newegg's site before leaping:

Thank you for your G.Skill purchase and review. Glad to hear the Trident Z memory kit is working well with your X99 system. DDR4-3200 is pushing the limits of the Haswell-E CPU, so additional tweaks may be necessary to stabilize the processor. Manually set the System Agent Voltage to see if that gives the memory controller the boost it needs to stabilize a higher DRAM frequency. Once the proper values are input, the computer should be perfectly stable. For any further questions or if you need additional help with set up, please feel free to contact us directly.

Thank you
GSKILL SUPPORT

Source

Keep in mind that this does not affect Z170 boards - they do not need an OC'd strap to use this RAM.

 

[Carfax83]

They were $225. A Newegg mobile coupon would have knocked it down another 5%.

Oh, I didn't miss much then

That's a pretty old thread. You might want to see this comment from G.skill on Newegg's site before leaping:

I think the guy that wrote that review made a mistake. He bought the dual channel kit, which is specifically for the Z170 boards rather than the quad channel kit, which is the one that I bought.

The dual channel kit has matched pairs, and more aggressive timings than the quad channel kit which presumably has matching quadruplets with more relaxed timings.

This is the set that I got. Notice how the timings are less aggressive than the set in the link you posted?

At any rate, if I can't run them at 3200MHz it's not a big deal. Like you said earlier, on a quad channel board the RAM speed isn't as much of a factor due to higher available bandwidth and the larger caches of the CPU..

5930K has almost twice the L3 cache size as the 6700K, so that means less trips to system memory for cache misses.. :thumbsup:

 

[Termie]

Oh, I didn't miss much then



I think the guy that wrote that review made a mistake. He bought the dual channel kit, which is specifically for the Z170 boards rather than the quad channel kit, which is the one that I bought.

The dual channel kit has matched pairs, and more aggressive timings than the quad channel kit which presumably has matching quadruplets with more relaxed timings.

This is the set that I got. Notice how the timings are less aggressive than the set in the link you posted?

At any rate, if I can't run them at 3200MHz it's not a big deal. Like you said earlier, on a quad channel board the RAM speed isn't as much of a factor due to higher available bandwidth and the larger caches of the CPU..

5930K has almost twice the L3 cache size as the 6700K, so that means less trips to system memory for cache misses.. :thumbsup:

Oh, good catch, it was actually the dual channel kit I was talking about being on sale - didn't realize it was targeted only at Z170. Not sure what yours was priced at.

Interesting if looser timings really allow a different strap. Please report back!

 

[Hail The Brain Slug]

When I built my X99 system a few weeks ago I got two of these.

I know 3000 isn't as hard to achieve as 3200 with the HW-E chips, but I have not had any issues with it running a 125MHz strap. However, I started working on overclocking my chip as soon as windows was running on it so if there were stability issues, I worked them out through my overclocking testing.

 

[Termie]

When I built my X99 system a few weeks ago I got two of these.

I know 3000 isn't as hard to achieve as 3200 with the HW-E chips, but I have not had any issues with it running a 125MHz strap. However, I started working on overclocking my chip as soon as windows was running on it so if there were stability issues, I worked them out through my overclocking testing.

There's no real issue hitting DDR4-3000, it's the fact that a 125MHz strap is necessary.

Carfax - I went ahead and tested DDR4-3200/16-18-18-38, and sure enough, it worked on a 100MHz strap. Very surprised, given that DDR4-3000/15-15-15-35 couldn't post with that strap.

But...I did some quick system benchmarks, and it was actually slower at the DDR4-3200 setting than my typical DDR4-2666/15-15-15-35. So, you might want to try both out. Could be that the weak timings at 3200MHz have an overly detrimental effect. Again, I'd appreciate any updates you can provide, as there isn't a lot of discussion of this issue out there.

 

[Carfax83]

When I built my X99 system a few weeks ago I got two of these.

I know 3000 isn't as hard to achieve as 3200 with the HW-E chips, but I have not had any issues with it running a 125MHz strap. However, I started working on overclocking my chip as soon as windows was running on it so if there were stability issues, I worked them out through my overclocking testing.

Why did you buy two dual channel kits? I would think that would make it more difficult to attain full stability seeing as they only come in pairs, unlike the quadruplets in a quad channel kit.

 

[Carfax83]

Carfax - I went ahead and tested DDR4-3200/16-18-18-38, and sure enough, it worked on a 100MHz strap. Very surprised, given that DDR4-3000/15-15-15-35 couldn't post with that strap.

Glad to see it worked! :thumbsup:

But...I did some quick system benchmarks, and it was actually slower at the DDR4-3200 setting than my typical DDR4-2666/15-15-15-35. So, you might want to try both out. Could be that the weak timings at 3200MHz have an overly detrimental effect. Again, I'd appreciate any updates you can provide, as there isn't a lot of discussion of this issue out there.

Thats weird. And your DDR4 2666 settings use 100 or 125 strap?

If it's using 125 strap, then the uncore frequency is going to be higher which would explain the performance increase, compared to running at 100 which is the default. At any rate, you could always try to tighten the settings a bit at 3200 to see if you can squeeze more performance out of it.

Did you run any latency benchmarks?

 

[Termie]

Glad to see it worked! :thumbsup:



Thats weird. And your DDR4 2666 settings use 100 or 125 strap?

If it's using 125 strap, then the uncore frequency is going to be higher which would explain the performance increase, compared to running at 100 which is the default. At any rate, you could always try to tighten the settings a bit at 3200 to see if you can squeeze more performance out of it.

Did you run any latency benchmarks?

DDR4-2666 has always worked for me at the 100MHz strap, and that's what I used (I think the thread you linked to said 2400MHz was the limit, but I haven't found that). I didn't test for latency - I just did some processor benchmarks.

You're right - tightening timings might help, although I was keeping in mind the potential issue with tight timings on X99 and just wanted to see if the strap worked...and it did.

 

[Carfax83]

DDR4-2666 has always worked for me at the 100MHz strap, and that's what I used (I think the thread you linked to said 2400MHz was the limit, but I haven't found that). I didn't test for latency - I just did some processor benchmarks.

Yeah, that link to the thread I posted earlier was fairly old, so the motherboard makers have been busy since then tweaking their BIOS..

At any rate, I think the actual problem of lower than expected performance is one of the peculiarities of the X99 platform. If the strap and uncore frequency are the same, then the DDR4 3200 should always be faster, as the frequency boost over the 2666 modules should more than make up for the slacker timings.

Was 1T command rate used for both?

On Skylake, you never really see this kind of divergence when it comes to memory performance. The faster memory is nearly always in the lead, which is because you don't need to tamper with the strap to run certain memory speeds..

Hardware canucks did an excellent review comparing DDR3 to DDR4 on Skylake, and the performance results are just about what you'd expect..

The DDR4 3200 leads in every single test I believe, and the latency lowers with increased RAM speed:

 

[Carfax83]

Perfect example. They did a review of the Corsair Dominators but this time with a Haswell-E setup, and sure enough their results are all over the place because of having to mess with the strap:

 

[Termie]

Perfect example. They did a review of the Corsair Dominators but this time with a Haswell-E setup, and sure enough their results are all over the place because of having to mess with the strap:

This was a heroic effort to compare different memory speeds, but you can see how difficult it is with the odd strap.

Note that in the system benchmarks, DDR4-3400 never won. That's kind of what I was getting at - don't just test latency, because it doesn't give an accurate picture of performance. In all of their full-system tests, DDR4-3000 to DDR4-3200 with tighter timings were faster.

In other words, you'll almost certainly be better off running your memory at something under 3200/16, but then you'll be forced into a bad strap situation. That's the trouble with Haswell-E's handling of memory speed.

 

[Carfax83]

That's the trouble with Haswell-E's handling of memory speed.

Yeah, the uncore running at different clock speeds makes it much harder to get an accurate comparison.. Even so, the fact that Haswell-E has such a large L3 cache and much more bandwidth to begin with also explains why the results don't scale as linearly as Skylake.

Skylake being dual channel and having less L3 cache makes it put more pressure on the system memory, and so you're going to see greater differences versus Haswell-E just due to that as well.

When I get all my parts in, the uncore is the last thing I'm going to fiddle with. Core speed is king when it comes to CPUs, and I think the uncore only makes a difference for synthetic benchmarks.

I'm aiming for 4.4ghz on the core, and I want to try to get the memory to 16-16-16-36 at 3200MHz using the 100 strap..

This is the best overclocking guide I could find for Haswell, and they go into whether or not overclocking the uncore is worth it or not.

It seems the consensus is no, it's not worth it..

Honestly, these memory tests don't come close to taxing or exploiting the bandwidth available to the Haswell-E platform.. The only consumer software that makes use of this bandwidth is compression software like 7zip and a small amount of games, none of which were tested here..

 

[tweakboy]

Dual SLI with 980 Ti 's

 

[Carfax83]

Well, I wasn't able to run at DDR4 3200 at XMP timings at my preferred settings (4.3GHz using 1.2v adaptive and cache at 3700GHz), let alone at 16-16-16-36..

Maybe if I fiddled with the timings and voltages more I could do it, but it's just not worth it....at least not right now. Trying to run high RAM speeds with high capacity DIMMS is more difficult than I thought it would be..

Funnily enough, I settled on the exact same speed and timings as you, DDR4 2666 at 15-15-15-35 CR1. At this speed, the RAM is fast enough so as not to be a bottleneck, and you can increase your CPU and cache speed without really having to worry about your RAM becoming unstable as a side effect..

Core speed is king after all :thumbsup: All in all, I'm very satisfied with my new build. It's noticeably faster than my previous x79 build, and my games feel smoother even at my GPU limited settings. And Windows 10 is definitely using all of the RAM, even if it's mostly to cache data..

 

[Termie]

Thanks for the update!

Yeah, RAM speed is definitely overrated, especially for those of us looking to run our CPUs at a big overclock. Strange that you couldn't get XMP to work even with sticks approved for X99 use. Must just be the interaction with the CPU overclock that made it impossible. You're running really low vcore, which is awesome for CPU temps, maybe not great for RAM OC stability.

 

[Carfax83]

Thanks for the update!

Yeah, RAM speed is definitely overrated, especially for those of us looking to run our CPUs at a big overclock. Strange that you couldn't get XMP to work even with sticks approved for X99 use. Must just be the interaction with the CPU overclock that made it impossible. You're running really low vcore, which is awesome for CPU temps, maybe not great for RAM OC stability.

Yeah the XMP settings worked fine till I started to overclock manually. Then it became unstable..

But like you said, it's very likely a voltage issue. 4.3ghz is the sweet spot for my CPU, as I'm able to run at that speed with perfect stability using a Vcore of 1.2v with an offset of 0.005v..

And then I have the cache running at 3.6ghz. I bet if I fiddled with the system agent voltage I could get 3.2ghz along with my other settings.

Don't have the will to do it now though.... but one day

 

[Termie]

FYI -

My 5820K is at 4.4GHz, with stock uncore (3.3GHz). I tried upping it all the way to 4GHz but saw no improvement in gaming performance, only an increase in power use.

I use an offset voltage. I believe it's hits 1.2V at load - I'm not at home right now so I can't quite recall. The memory is 4x8GB of DDR4-2666 at 15-15-15-35-1N, 1.2V.

 

[Carfax83]

FYI -

My 5820K is at 4.4GHz, with stock uncore (3.3GHz). I tried upping it all the way to 4GHz but saw no improvement in gaming performance, only an increase in power use.

I use an offset voltage. I believe it's hits 1.2V at load - I'm not at home right now so I can't quite recall. The memory is 4x8GB of DDR4-2666 at 15-15-15-35-1N, 1.2V.

Our rigs are very similar then. I might try offset later to see whether that improves my stability when running DDR4 3200, because it's not stable right now while the CPU is overclocked.

I look forward to seeing your tests for DDR4 3200 vs DDR4 2666.

 

[Carfax83]

OK, I managed to squeeze out another 100mhz out of the CPU without increasing the voltage, by correcting a mistake I had with the settings.

So now it's 4.4ghz at 1.2v plus the 0.005v for offset. Basically I had it at minus before and not plus :$

Despite my best efforts though, my PC will not run at DDR4 3200....even at stock settings!

Which is really weird, because I installed Windows 10 with the XMP settings set to DDR4 3200. And I booted into Windows perfectly several times after the installation completed.

The only thing I can remember doing which might explain my current difficulties, is that I upgraded the BIOS revision from 1901 to 2001. But I switched back just a few hours ago to see if I could run XMP stable, and it didn't take :\

Anyway, I'm still highly pleased with my final settings. More than fast enough for everything I throw at it.. :thumbsup: