Should I go for faster speed or lower latency?

[UncleGrandpa]

Hi all, I'm pretty sure I will be going with the Extreme6 motherboard, 6600k and H7 (might upgrade later). But I can't figure out what memory to get. I'm especially confused by the latencies.

I know CL 15 is supposed to be faster than 16 or 17 and I read somewhere that a 3000 CL15 should be equivalent to a 3200 CL16 in terms of speed. Is that true? I'm having trouble deciding what to focus on.

Should I go with higher speed like 3200 even if latency is high, or should I go with lower latency even if speed is less?

Also, could someone explain this? I was just looking at 2 different sets of DDR4 3200. One was 16.16.16.36 and the other 16.18.18.38. Both are listed as CL16. Is one of these better than the other? Thanks so much.

If anyone wants to suggest a brand I'd really appreciate it.

 

[wingman99]

G.SKill is a good brand.

Here is a chart with cas lantncy and Memory speed. https://en.wikipedia.org/wiki/CAS_latency

 

[Woomack]

Theoretical latency/fequency has not much to do with real performance gains on new platforms. All last Intel series are showing higher performance gain from memory frequency than latency just because big part of access time is CPU's cache. When you overclock CPU cache then in most cases you see higher memory bandwidth/latency gain than from higher memory "speed". There are couple of variables so not always can tell what is faster as all depends on application which you are using.
In general it's always like large files work better on higher memory bandwidth. A lot of smaller files work better at lower latency. There is also single/dual rank factor which is never considered in any memory performance guides/tables.

I recommend at least something like 3000 14-14-14/15-15-15 or 3200 15-15-15/16-16-16. G.Skill is manufacturing all their Trident Z and most ( if not all ) Ripjaws V on Samsungs which are the best for overclocking right now. Other brands are using mixed chips or only Hynix/Micron.
It doesn't mean this memory will give you much higher performance but it doesn't cost much more than "standard" memory kits. Wingman has one of these better kits in the sig

 

[mackerel]

There is also single/dual rank factor which is never considered in any memory performance guides/tables.

This! I really wish I understood it more myself, but in my areas of interest I find dual rank gives about 20% boost in application performance over single rank if all else is equal. It doesn't show up in synthetic bandwidth tests or things like that. So for a high performance system I'd use that 20% as a guideline to how much faster SR would have to be clocked over DR to be worth it. Note the modules don't have to be DR themselves, as two SR modules per channel (4 modules in dual channel) give the same benefit. At least in my use cases, latency gives an insignificant difference compared to clocks, so I don't worry about them.

I went a step further and tried looking to see if this could impact gaming also, where it might get more attention. I didn't succeed in that, without significant framerate differences in Witcher 3 and Doom which I used for test. Maybe I need more "CPU" demanding games which are said to get the most benefit. I got GTA5 on sale since then so maybe that'll be a better choice. Actually, with hindsight maybe I should have lowered the graphical quality to push the balance more to CPU... but certainly in sensible settings at the time it didn't help.

 

[EarthDog]

Great information listed here.

Now comes the great memory discounter...

Sweetspot is ddr4 3000-3200 cl14. I'd stick right around there and not got lost in the memory minutia as very few things improve notably, if it all.

 

[UncleGrandpa]

Thanks so much everyone I've narrowed down to a couple of brands/models. I was almost going to buy a Geil 3200 that was 16-16-16 but I put it in my shopping cart and then came back 20 minutes later and they were all gone, like sold out Such bad luck in that 20 min gap lol.

I was given a 10% off coupon today by newegg and I'm wondering if I should get a Ripjaws V 3200? The problem is it's not 16-16-16, it's actually 16-18-18. Is that bad? All things being equal, are the 16-16-16 much better?

Thanks so much for the suggestion Earthdog, I would totally buy a CL14 it's just that they are mostly outside my budget. I'm trying to stick under $100 no matter what (if it's under $85 even better) haha.

 

[EarthDog]

http://www.newegg.com/Product/Produ...&cm_re=2x8gb_ddr4_3000-_-20-158-070-_-Product

 

[UncleGrandpa]

http://www.newegg.com/Product/Produ...&cm_re=2x8gb_ddr4_3000-_-20-158-070-_-Product

I was just looking at this exact set and trying to figure out if I should go with it or the Ripjaws V 3200 that's also $85

This one is timing 15-17-17-35 for 3000 frquency and the Ripjaws V is 16-18-18-38 for 3200. Which one would you say is better? I'm fine with anything over 3000, just having a hard time figuring out which one's a better product since they cost the same right now

This is the GSkill http://www.newegg.com/Product/Product.aspx?Item=N82E16820232181&ignorebbr=1

 

[wingman99]

I was just looking at this exact set and trying to figure out if I should go with it or the Ripjaws V 3200 that's also $85

This one is timing 15-17-17-35 for 3000 frquency and the Ripjaws V is 16-18-18-38 for 3200. Which one would you say is better? I'm fine with anything over 3000, just having a hard time figuring out which one's a better product since they cost the same right now

This is the GSkill http://www.newegg.com/Product/Product.aspx?Item=N82E16820232181&ignorebbr=1

They come out the same for latency in access time (ns), however ram speed helps in pipelining.

So I would go with the 3200 speed.

With asynchronous DRAM, the time delay between presenting a column address and receiving the data on the output pins is constant. Synchronous DRAM, however, has a CAS latency that is dependent upon the clock rate. Accordingly, the CAS latency of an SDRAM memory module is specified in clock ticks instead of absolute time.

Because memory modules have multiple internal banks, and data can be output from one during access latency for another, the output pins can be kept 100% busy regardless of the CAS latency through pipelining; the maximum attainable bandwidth is determined solely by the clock speed. Unfortunately, this maximum bandwidth can only be attained if the address of the data to be read is known long enough in advance; if the address of the data being accessed is not predictable, pipeline stalls can occur, resulting in a loss of bandwidth. For a completely unknown memory access (AKA Random access), the relevant latency is the time to close any open row, plus the time to open the desired row, followed by the CAS latency to read data from it. Due to spatial locality, however, it is common to access several words in the same row. In this case, the CAS latency alone determines the elapsed time.
Formulas used to calculate latencies
Jump up ^ Bit time = 1/Data rate
Jump up ^ Command rate = Data rate/2
Jump up ^ Cycle time = 2*Bit time
Jump up ^ First word = ( (CAS latency*2) + (1-1) ) * Bit time
Jump up ^ Fourth word = ( (CAS latency*2) + (4-1) ) * Bit time
Jump up ^ Eighth word = ( (CAS latency*2) + (8-1) ) * Bit time https://en.wikipedia.org/wiki/CAS_latency

 

[UncleGrandpa]

Thanks Wingman! I think I will settle for that one if they will honor my 10% coupon lol. Just curious though, are the TridentZ much better in comparison to these? I've seen lots of people recommending them.

 

[wingman99]

The memory you are comparing use the same IC's under the heat spreaders as the TridentZ. Memory is very simple to compare performance, just the speed and CL timings is all folks need to compare.

 

[Woomack]

Actually Ripjaws V and Trident Z 2x8GB 3200 16-18-18 can be on single or dual rank IC. I was reviewing Trident Z 3200 16-16-16 and 3400 16-18-18 and both were on dual rank Samsungs. 2x8GB 3000 14-14-14 and 3200 14-14-14/15-15-15 are on single rank. Ranks are affecting performance and overclocking potential. Usually dual rank is faster but is overclocking worse. There are other factors which are affecting performance so can't just look at the CL and tell what is better. On Skylake platform higher frequency is giving more than tighter timings. Best if both at the same time but there are some limits.

 

[UncleGrandpa]

Hi Woomack, thanks for sharing. I'm a bit clueless about all this (majored in a social science lol) but I'm hoping to build a good gaming PC for my son so I'm reading as much as I can. How can I tell if a stick is dual or single rank? I have the Extreme6, not sure if that makes a difference in what kind/brand of ram might work better for me.

I still haven't made my decision, figured I could wait a bit more because I won't be getting the 6600k until Feb so I'd just have the memory lying around the house anyway. But I've narrowed down to the Geil Forza 3200 16-16-16 that went out of stock, G.Skill Ripjaws V 16-18-18, and Corsair LPX which is just $80 right now for the 3000 15-17-17 sticks. Is there any of these you'd recommend over the others?

I see "Team Group" memory on sale all the time but I'm scared to buy it because I never heard of this brand before

 

[wingman99]

Actually Ripjaws V and Trident Z 2x8GB 3200 16-18-18 can be on single or dual rank IC. I was reviewing Trident Z 3200 16-16-16 and 3400 16-18-18 and both were on dual rank Samsungs. 2x8GB 3000 14-14-14 and 3200 14-14-14/15-15-15 are on single rank. Ranks are affecting performance and overclocking potential. Usually dual rank is faster but is overclocking worse. There are other factors which are affecting performance so can't just look at the CL and tell what is better. On Skylake platform higher frequency is giving more than tighter timings. Best if both at the same time but there are some limits.

Dual rank is slower performance also faster perfromance depending on the pipeline stall, data and memory controller.

Performance of multiple rank modules[edit]
There are several issues to consider regarding memory performance in multi-rank configurations.

Multi-rank modules allow several open pages in each rank (typically 8 pages per rank). This increases the possibility of getting a hit on an already open row address. The performance gain that can be achieved is highly dependent on the application and the memory controller's ability to take advantage of open pages.[citation needed]

Multi-rank modules have higher loading on the data bus (and on unbuffered DIMMs the CA bus as well). Dual rank DDR3 DIMMs can run at DDR3-1600, but if there are more ranks connected in one channel, the speed will be reduced.[citation needed]

Subject to some limitations, ranks can be accessed independently, although not simultaneously as the data lines are still shared between ranks on a channel. For example, the controller can send write data to one rank while it awaits read data previously selected from another rank. While the write data is consumed from the data bus, the other rank could perform read-related operations such as the activation of a row or internal transfer of the data to the output drivers. Once the CA bus is free from noise from the previous read, the DRAM can drive out the read data. Controlling interleaved accesses like so is done by the memory controller.[citation needed]

There is a small performance reduction for multi-rank systems as they require some pipeline stalls between accessing different ranks. For two ranks on a single DIMM it might not even be required, but this parameter is often programmed independently of the rank location in the system (if on the same DIMM or different DIMMs). Nevertheless, this pipeline stall is negligible compared to the aforementioned effects.[citation needed]
https://en.wikipedia.org/wiki/Memory_rank

 

[dudleycpa]

Actually Ripjaws V and Trident Z 2x8GB 3200 16-18-18 can be on single or dual rank IC. I was reviewing Trident Z 3200 16-16-16 and 3400 16-18-18 and both were on dual rank Samsungs. 2x8GB 3000 14-14-14 and 3200 14-14-14/15-15-15 are on single rank. Ranks are affecting performance and overclocking potential. Usually dual rank is faster but is overclocking worse. There are other factors which are affecting performance so can't just look at the CL and tell what is better. On Skylake platform higher frequency is giving more than tighter timings. Best if both at the same time but there are some limits.

If I understand you Sir Woomack something like this? Mind you it isn't $85.

G.SKILL Ripjaws V Series 16GB DDR4 3200 $125
http://www.newegg.com/Product/Product.aspx?item=N82E16820232229

 

[Woomack]

Dual rank is slower performance also faster perfromance depending on the pipeline stall, data and memory controller.

There are no performance drops on current gen desktop chipsets using dual rank memory. You can probably see it on some 4 rank+ platforms so servers. I've never seen worse performance using dual rank DDR3 or DDR4 on desktops while I've seen it on 8 rank server setup.
On all Intel 6 series+ chipsets there is no worse performance using dual rank modules. Some applications are showing higher performance and sometimes it's really visible as for example HyperPi 32M can sometimes run as fast on 2400 dual rank as on 3000 single rank modules.

The best DDR3 IC for competitive benchmarking was dual rank - PSC, Hypers, Samsung HYK0 etc.
The best DDR4 IC for overclocking is Samsung B just because you can OC it above 4000 at tight timings so frequency and latency is beating difference made by ranks. Dual rank Samsung B so in 16GB modules can't really OC past 3600 and can't run at so tight timings above ~3400 as single rank modules.

If I understand you Sir Woomack something like this? Mind you it isn't $85.

G.SKILL Ripjaws V Series 16GB DDR4 3200 $125
http://www.newegg.com/Product/Product.aspx?item=N82E16820232229

Yes, it's one of the kits which is probably on Samsung B single sided/single rank IC.


Here is what Crucial says about clock vs latency:
http://www.crucial.com/usa/en/memory-performance-speed-latency

 

[mackerel]

Some applications are showing higher performance and sometimes it's really visible as for example HyperPi 32M can sometimes run as fast on 2400 dual rank as on 3000 single rank modules.

Prime95 and similar prime number finding software based on the same library are another example I have come across. Where large FFT sizes are used, it is easily ram bandwidth limited e.g. using a fast quad core CPU. Having the same speed ram in dual rank vs. single rank gives about 20% performance improvement. I remember comparing benchmarks once with someone with highly overclocked 6700k and 3700 ram, which I beat with mine at only 4.2 GHz and 3200 dual rank ram.

 

[Kenrou]

G.SKILL Ripjaws V Series 16GB DDR4 3200 $125
http://www.newegg.com/Product/Product.aspx?item=N82E16820232229

Easy and stable OC to 3600mhz 15-15-32 1.4v, 3733mhz 15-15-34 1.45v but motherboard stopped me from going above, likely get 4000mhz+ from ASRock or MSI

 

[dudleycpa]

Easy and stable OC to 3600mhz 15-15-32 1.4v, 3733mhz 15-15-34 1.45v but motherboard stopped me from going above, likely get 4000mhz+ from ASRock or MSI

That certainly is worth $40 in the price differential department (IMHO).

Anyone notice how suddenly motherboard prices & RAM prices have spiked?

 

[Woomack]

RAM costs more because of high demand in mobile devices and brands like Samsung switched production. I haven't noticed difference in motherboard prices.
I wanted to buy one memory kit but prices in my area went up by 20-30% and it's too much for already overpriced kits which I usually buy.