Those of you who can calculate CL/MHZ in here please!

[spat55]

Which is fastest, CL18/3600mhz or CL16/3200mhz? I don't quite understand the way I'm meant to calculate it.

[Megah3rtz]

cl16 3200mhz 

[LtStaffel]

 

[MageTank]

14 minutes ago, spat55 said:

Which is fastest, CL18/3600mhz or CL16/3200mhz? I don't quite understand the way I'm meant to calculate it.

Most people do CL / MHZ x 2000. So in the 3200mhz example, it would be 16 / 3200 x 2000 = 10ns (absolute latency, in clock cycles).

 

Now, this can be pretty deceptive in the grand scheme of things, as this number only represents a "potential" speed. Real world, far too many factors exist that could allow for ram with lower frequency or looser primary timings to be faster. For example: A kit that is 3000 C15 (10ns) can be faster than a kit that is 3200 C16 (also 10ns) because it could have tighter secondary and tertiary timings that are not even advertised. 

 

In the grand scheme of things, the only thing that matters is bandwidth efficiency, and RTL (Round Trip Latency). While clock speed and primary timings do play a pretty significant role in determining your RTL value, it means nothing if you have to sacrifice on the efficiency of other timings in order to achieve a higher clock, something you will notice when comparing extremely high frequency XMP's against properly tuned manual overclocks at much lower frequencies. 

 

While a higher frequency kit has a much higher theoretical bandwidth (sometimes referred to as a bandwidth ceiling), it has a much more difficult time getting anywhere near that theoretical number, as you tend to require much looser tertiary timings to facilitate that frequency. 

 

Hopefully this makes sense, and I am sorry if I over-complicated a simple question. I merely want to point out that it's a little more complicated than just comparing advertised numbers on a marketing sheet. Not all ram, memory controllers, and motherboards are made equal. 

[Jade]

While those two posts above are correct, @MageTank has a some good explanations in his overclocking memory guide.

 

And as I was about to post this, here he was.

[MageTank]

1 minute ago, Jade said:

While those two posts above are correct, @MageTank has a some good explanations in his overclocking memory guide.

 

And as I was about to post this, here he was.

I do go into specifics regarding bandwidth efficiency in the guide, but a lot of people seem to get confused when I try to explain that portion, so I try to avoid it when necessary, especially once numbers get involved. Telling someone "you have 90% write efficiency" seems to throw them for a loop, even with the best of my explanations.

 

To try to simplify it again, when I say "x% efficiency", it means you are achieving that percentage of your peak theoretical bandwidth. Using 3200mhz as yet another example, it has a peak theoretical bandwidth of 51.2GB/s in dual channel (3200 x 8 x 2). If you are getting 46GB/s write bandwidth, you are achieving 90% write efficiency. If you are getting 43.5GB/s, you are achieving 85% read efficiency.

 

As complicated as this is to grasp for some people, I cannot stress the importance of this information enough. Some might believe they are getting faster ram if they go for a higher clocked kit, but if their efficiency is poor, it's going to perform identically (and sometimes, worse) than a slower kit. A 3200 kit with 90% efficiency is going to perform worse than a 3000 kit with 96% efficiency. A higher frequency only helps if you can maintain high efficiency alongside it. Simply raising your frequency does nothing for you if you have to compromise on timings that directly impact efficiency. 

 

Again, I feel like I am over-complicating this, so I'll let it rest at that. 

[spat55]

2 minutes ago, MageTank said:

I do go into specifics regarding bandwidth efficiency in the guide, but a lot of people seem to get confused when I try to explain that portion, so I try to avoid it when necessary, especially once numbers get involved. Telling someone "you have 90% write efficiency" seems to throw them for a loop, even with the best of my explanations.

 

To try to simplify it again, when I say "x% efficiency", it means you are achieving that percentage of your peak theoretical bandwidth. Using 3200mhz as yet another example, it has a peak theoretical bandwidth of 51.2GB/s in dual channel (3200 x 8 x 2). If you are getting 46GB/s write bandwidth, you are achieving 90% write efficiency. If you are getting 43.5GB/s, you are achieving 85% read efficiency.

 

As complicated as this is to grasp for some people, I cannot stress the importance of this information enough. Some might believe they are getting faster ram if they go for a higher clocked kit, but if their efficiency is poor, it's going to perform identically (and sometimes, worse) than a slower kit. A 3200 kit with 90% efficiency is going to perform worse than a 3000 kit with 96% efficiency. A higher frequency only helps if you can maintain high efficiency alongside it. Simply raising your frequency does nothing for you if you have to compromise on timings that directly impact efficiency. 

 

Again, I feel like I am over-complicating this, so I'll let it rest at that. 

No that actually makes sense, you are saying that some kits with slower frequency are faster because they can hit their maximum efficiency closer (on average) with tighter timings unlike faster kits which can't get as close. 

[MageTank]

5 minutes ago, spat55 said:

No that actually makes sense, you are saying that some kits with slower frequency are faster because they can hit their maximum efficiency closer (on average) with tighter timings unlike faster kits which can't get as close. 

Exactly! As frequency gets higher, it puts more stress on not only the IMC, but the traces on the board, resulting in poor (extremely loose, and even unstable) training for the rest of your timings. Due to this poor training, your faster kit can actually be magnitudes slower. I've seen kits have efficiency as poor as 70% before, and get nowhere near their theoretical bandwidth.

 

That's not an excuse to invest in slow memory on purpose, but it's certainly something to keep in mind when spending more money on a "faster" kit. I advise people to invest in quality IC's if they can, even if it's faster than what they feel they need, and instead "underclock" the kit while tightening it as much as possible. I know people that buy 3600 C16 kits just to tighten them down to 3200 C12 and end up with faster ram as a result, while also reducing stress on both the board and IMC at the same time. I know people that run extremely tight kits with 0.950v VCCIO/VCCSA that are faster in every regard than people that need 1.2v VCCIO/VCCSA just to maintain their kits. That's generally how you know you've found the perfect memory OC, when you can get more out of it for less investment. I was told this by the man that taught me how to OC memory, and it still holds true after all these years. 

[johndms]

@MageTank Not to hijack someone's thread, but it's sorta on topic. What's your opinion on Ryzen system calculator v 0.9.5 + Records? I saved that to my Google Drive and used it on my 3000MHz Hynix M-Die kit. Instead of tightening the usual CL, tRCD, tRP, tRAS, tRC.. I've begun experimenting with the tRFC Calculator (tab at bottom) and lowering those numbers while leaving my primary timings at default. Is this what you're referring to above when you mention tighter secondary and tertiary timings?

 

Edit: I just went back and looked at your guide again, perhaps I should pay closer attention to what I read.

[MageTank]

6 minutes ago, johndms said:

@MageTank Not to hijack someone's thread, but it's sorta on topic. What's your opinion on Ryzen system calculator v 0.9.5 + Records? I saved that to my Google Drive and used it on my 3000MHz Hynix M-Die kit. Instead of tightening the usual CL, tRCD, tRP, tRAS, tRC.. I've begun experimenting with the tRFC Calculator (tab at bottom) and lowering those numbers while leaving my primary timings at default. Is this what you're referring to above when you mention tighter secondary and tertiary timings?

First of all, thank you for that page. I'll certainly study it once I get more time on my hands, but as far as tRFC goes on Ryzen, it's quite enigmatic to me. Unlike Intel, Ryzen has multiple tRFC values. Not only that, but in my personal testing, tRFC2 and tRFC3 seemed to do absolutely nothing on the Ryzen 1600 I was testing with, so I was not able to figure out much about it.

 

As for secondary timings, tRFC is indeed a secondary timing. Your primary timings tend to dictate the absolute potential of your ram, while secondary/tertiary timings determine exactly how close you can get to that potential.Gauging the total speed of ones ram by using frequency and primary timings alone would only work in a perfect world, so you can see why it's pretty important to focus on efficiency over trying to "raise the ceiling" . 

 

Now that I have my dream job, I'll be able to allocate enough funds to get my own personal ryzen test bench, instead of only being able to tinker with my friends system. I want to buy the cheapest CPU/board I can get, and then test various qualities of memory so that I can make safe recommendations to people. My logic being, if the worst CPU's and boards can scale well with X kinds of memory, the better boards/CPU's should be able to do so as well. I find it to be kinda misleading if I just throw my money at a high end board/CPU and say "you should be able to do this on your $80 motherboard, because my $200+ board can do it just fine". I've noticed this when I tried to compare the superior trace topology of my expensive ITX board, and the Apex/OCF to people using cheaper Intel boards, and wondering why they were unable to achieve similar results.

 

I will say this though, some of the results in the document you linked are quite impressive. My personal attempt at taming latency on Ryzen achieved 76ns on latency. These people managed to get down to 65ns. While that is still extremely high as far as DDR4 latency goes, it's still impressive they managed to get that far given the limitations of the IMC. Hopefully I can break into the 50's, even if it's 59.9ns, lol.