How important are timings with DDR4?

[ssjwizard]

So I know that as generations of RAM have progressed the importance of tight timings has become more and more of a non issue. I am potentially looking to upgrade soon and I find myself more concerned with the look of the heat spreaders than I do with the actual specs. Anyways so lets take this to a benching and tweaking level for a moment. The root question here is how relevant are tighter timings on DDR4? I mean is there any appreciable difference between CL 13, 14, 16 memory at this stage? Im not looking for a purely technical answer nor a day to day use because I already know the answer to those questions. What Im interested in is in high demand situations does buying a CL 13 kit over a CL 16 kit really matter anymore?

Secondary question outside of benchmarking and epeen does higher speed memory even mater now?

 

[Woomack]

For gaming anything ~3200 is good if you look at performance to price factor. Large and fast CPU cache is covering any latency delays so regardless if memory is CL13 or CL16 it won't really matter out of benchmarking.

For benching best is anything on 2x8GB single rank samsungs so new TridentZ 3000/3200 14-14-14, 3600 16-16-16 or something similar.

If we are talking about general performance and how timings are affecting it then I will say that higher frequency > tighter timings.
However if you are looking for cheaper motherboard then you will be limited to ~3200-3466 memory frequency.

At the end you don't have big choice. There are older Hynix kits and newer Samsung kits. Hynix in mass sales is almost only available on 1 IC known as MFR. Samsung is available on 3-4 IC but can divide them to older - 2x4 single sided/2x8GB double sided and newer 2x8 single/2x16GB double sided.

Regardless what frequency memory kit you buy it will be something like below:

If you get Hynix then you will end at:
- ~3200 15-15-15/16-16-16 1.35V

If you get older Samsung then you will end at:
- ~3466 16-18-18/17-18-18 1.35-1.45V - double sided kits
- ~3733 18-19-19 1.35-1.45V - single sided kits

If you get newer Samsung then you will end at:
- ~4000 18-18-18 1.35V - single sided ( so 2x8GB )
I have no idea how are overclocking 2x16GB kits

There are also Micron kits but barely anyone is using them. You can find them almost only in Crucial kits which in general won't make much more than 3000. Some can make really tight timings at low voltages like 3000 12-12-12 1.40V.

So to make high frequency you need good memory but also what is more important, good motherboard.
Good overclocking memory kits can cost about as much as these worse. The same with motherboards so simply if you want to buy something not overpriced then read some about products you are interested in.

I'm testing this kit right now:
http://www.newegg.com/Product/Produ...01&cm_re=tridentz_3000-_-20-232-201-_-Product
I think it's in reasonable price, looks great and overclocks like 3600 16-16-16 kit.

 

[mackerel]

One of my personal interest is prime finding, and that could be a "high demand" situation. I've long been looking into what influences the performance limitations of the software. My understanding has increased significantly over the last two weeks or so while I've been trying to figure out an unexplained massive performance difference between two otherwise very similar systems. In short, it was the ram. It is known that for bigger tasks that wont fit within the processor cache, ram performance (or lack thereof) will impact the scaling. Run one task, and it is practically unlimited in performance. Run 2, you usually get double the throughput. Run 3, this is when you tend to see some impact getting perhaps 2.7x the performance relative to 1x. Run 4, and you'll struggle to get much more than 3x. Getting my first Skylake system with high speed ram was an eye opener. Over at PrimeGrid, I had one of the fastest systems when running 4 large tasks, thanks to the fast ram. Testing showed I got tens of % speed improvement from running ram at 3333 compared to standard 2133. It wasn't entirely proportional to the ram clock, but still a very nice increase.

Long story short, I tried to replicate this in another system, but no matter what I did, it always ran some 25% slower. Eventually I found out if I had 4 sticks of ram in not 2 (same speed, timing), performance was where I expected it to be. As it is only a dual channel system, this isn't from going quad channel, but there may be some other optimisation in the memory system from having 4 modules present.

Looking through the test results I had, which included a little playing with timing, I saw 0-3% increase from running at 3000 16-16-16-36 compared to 16-18-18-38 so not really significant and might just be measurement variation. I didn't try tighter CAS. Running at 3333 compared to 3000 at the same timing values showed a 10% or so increase, which would be in line with the clock difference.

So at least in this application, bandwidth is particularly important, up to the point where there is enough and more doesn't affect it. I don't have an exact value but I'd expect overclocked quad channel ram to be practically unlimited. I don't (yet) have a quad channel ram system to play with to test that theory, as the high cost of that system means I could get better throughput from having two cheaper systems. Still, I see good scaling with ram speed on dual channel Skylake.

I recognise this is a very specific application which wont necessarily apply elsewhere.

For benching best is anything on 2x8GB single rank samsungs so new TridentZ 3000/3200 14-14-14, 3600 16-16-16 or something similar.

Is there an easy way to determine the rank of a particular module? Specifically do you know what rank is F4-3200C16D-16GVK tested previously? I'm still wondering if rank is in some way is influencing my observations elsewhere.

 

[Woomack]

In non-ecc DDR4 memory kits, if memory chips are on one side of PCB then it's single rank. If chips are on both sides then is dual rank. You can see it without removing heatsinks.

F4-3200C16D-16GVK was dual rank. I'm not sure if the same IC is in new kits under this product number.

 

[mackerel]

In trying to read up generally on rank, I understand it is related to the way the data is organised on module, which isn't necessarily related to the physical layout. Is it just so happens the sides = rank works on ordinary DDR4? I'm still trying to understand if higher overall rank system might provide a benefit as it is the only thing I have to go on right now with my 2 vs. 4 module results with other ram. The above kit is one I know that does give the higher performance in my use. I need to buy more ram but I need to be sure it will be a higher performance configuration before I do so.

 

[Woomack]

DDR4 has wider "internal bus" than DDR3 so ranks shouldn't matter in desktop PC. ECC memory can have more ranks and then it's hard to say how many ranks are on a module just looking at the memory.

In theory dual rank will be faster but it depends from some additional factors. Additionally 2x single rank modules in one channel should perform about as good as 1x dual rank module.
In DDR3 we could see that difference, in DDR4 not really. At least you can check some of my tests where results on single and dual rank memory are about the same. Difference ~1GB/s is really nothing. At the same time single rank modules are overclocking better.

I had a chance to compare performance on mentioned by you Ripjaws V 3200 16-16-16 memory in 2x4GB and 2x8GB. Both kits were about the same except capacity. Both were performing about the same too.

One more test which you can check is winsat so internal windows test. Run command prompt ( cmd ) and type winsat mem. It will perform memory tests.

 

[mackerel]

In the limited testing I've done so far comparing 2 vs 4 modules, I agree that the memory benchmarks I've run on it suggests 4 modules has slightly higher bandwidth, and slightly higher latency. This seems consistent with what you describe. I can also repeatably demonstrate Prime95 benchmark for core scaling shows a big (20%?) difference only changing between 2 and 4 modules. Those modules, I don't have the exact part on me, but they're Ripjaws 4 3333 ending GRRD, 4x4GB kit. The Ripjaws 5 kit above I also have, and they're a "fast" one. Prime95 benchmark is the only test I've found to show this effect, other than running long real tasks. At least, I haven't seen it in other CPU/memory benchmarks.

Comparing all the sub timings is still on the to do list, so maybe that could be a factor. The rank number is another area to look at.

I note on screenshots elsewhere Aida64 seems to be able to report rank. I'll see if I haven't tried the demo on one of the DDR4 machines but I might just give in and buy it after all.

 

[mackerel]

I had forgotten about Thaiphoon Burner and it appears that can also read the rank structure of a module. It confirms F4-3200C16-8GVK as dual rank, and my other modules F4-3333C16-4GRRD as single rank. So it seems to fit. Dual rank per channel = higher performance, single rank per channel = lower performance. In Prime95 anyway.

I'll probably play about a bit more tomorrow and see what happens with latency, and relative performance as the balance is shifted between CPU and RAM and see if I can work out what the minimum required bandwidth per core is for optimal performance. Then I'll see how stable it is as fast as I can get it before running real work through it.

Think I'll be hunting through Woomack's previous posts for tips to get that bit more speed out of ram... sure I saw it before, but can I find it again?

 

[ssjwizard]

This is pretty much what I was thinking. Im looking at the GEIL Dragon series ATM mostly because I find them visually appealing. Just need to decide if I want to go with the 2x8 or 2x16gb kit.