memory latency and speed numbers

[soulJAHmon]

OK, which brainiac/techy FA members can tell me about the memory numbers?

 

Are big numbers better ? or smaller numbers? Example: is 5-5-5-18 better or worse than 4-4-4-12?

 

I don't really care exactly what the numbers mean, all I want to know is am I looking for smaller or larger numbers, AND if smaller or larger numbers make a small or large difference in real-world performance.

[DJ aka GDR DJ]

the numbers are the timing from RAM

 

CAS latency (CL), clock cycle time (tCK), row cycle time (tRC), refresh row cycle time (tRFC), row active time (tRAS).

 

if you have not the ultra high, overclocked and i owned you all PC, have you the same (more or less) performance with low or high latency RAM

[soulJAHmon]

ok, so then unless you split hairs, and have very high performing pc, then higher or lower timing numbers doesn't matter much? Is that what you mean?

[rolf]

Those latencies describes how long you have to wait after asking the data of the memory. The number is measured in clock cycles on memory level.

 

Simple example, let's have DDRx memory at 1000MHz. Then you can take 1ns to send each data block. The most common latency is called the CAS latency (first number), which is the access time to access a column in the memory. A CAS latency of 5 means that you have to wait 5 cycles before you can read the data from that column, resulting in a delay of 5ns. Note that this latency if for a memory module! Hence, it is for single channel. If you have double channel DDRx, you still have for (both) modules the time to wait.

 

 

Of course, a delay of 4ns is better. If that's worth the extra money is an interesting question, but depends among others on the applications. For some applications, the difference is noticeable, for others it isn't. So let's use a nice analogy, a highway. Let the memory clock cycle be the speed you are allowed to drive, latency the distance you have (at most!) between the cars (in number of cars) and single/double/triple/quad channel the number of lanes. If you use only a few data (i.e. one car), latency and lanes doesn't matter, only speed. If you want to transport half of the US citizens to Peru, you'd better increase the number of lanes. If you request very often different columns, i.e. you randomly pick people from all states rather than first export San Diego, then LA, then Phoenix, etc, you have to switch columns often and hence the cas latency becomes important.

 

 

In the last case, your program sucks btw.  But it's apparently developers are allowed to ignore it In most applications this does not occur though.

 

 

The most important thing in every pc is not the memory clock, latency or number of channels. It is that you have sufficient memory. Because if you don't have enough, people can't use the highway but have to walk, i.e. use the swap space on the hard disk which is slow. To compare with respect to the analogy of the highway, if cars are driving with 120km/h or 70mph, people have to crawl backwards using only their left hand. That is a difference you want to avoid.

 

 

Last thing to say, you can overclock memory by in the bios setting the latency to some value lower than specified I've used that quite often, most memory modules which run at, e.g. 400MHz with CAS 5, can also run at 333MHz with CAS 4. If you run at 333MHz, you can buy slightly faster memory (the 400MHz) and set it manually to a lower value. Sometimes this is cheaper, since for 400MHz there exists 'value' ram which can run 'slower' with nice latency, while for 'slower' ram at 333MHz memory with specified low latency are 'extreme' ram. On the other hand, you'd have to manually configure this in the bios.

 

 

 

Personally I would buy a shitload of memory and ignore the latencies or speed. I'm at 24GB now

[Vanaraud]

As I understand it and the benchmarks shows also that sweetspot is 1600(GT\s?) CAS9 memory at moment. Tuning it down or upwards gives you minimal performance boost.

You have to check the latencies between same speed memory modules. 1600 memory with lower values could be slower from 2800 memory modules with higher values. As latencies are given in cycles and modules with higher speed can "run" through more cycles faster than with lower speed modules.

Personally if the PC is for multimedia: gaming, movies I don´t see the need for more than 8GB of memory. If one starts rendering videos\photos then the system actually uses16\24 memory.

[putts]

Personally if the PC is for multimedia: gaming, movies I don´t see the need for more than 8GB of memory. If one starts rendering videos\photos then the system actually uses16\24 memory.

What about 2 of those at the same time? Watching Netflix and playing a game at the same time? Heck, lets throw in encoding a video too.

[rolf]

As I understand it and the benchmarks shows also that sweetspot is 1600(GT\s?) CAS9 memory at moment. Tuning it down or upwards gives you minimal performance boost.

You have to check the latencies between same speed memory modules. 1600 memory with lower values could be slower from 2800 memory modules with higher values. As latencies are given in cycles and modules with higher speed can "run" through more cycles faster than with lower speed modules.

Personally if the PC is for multimedia: gaming, movies I don´t see the need for more than 8GB of memory. If one starts rendering videos\photos then the system actually uses16\24 memory.

 

The sweetspot depends on more than "at the moment". An i3 has sufficient with less memory performance than an i7, it can't use everything. On the other hand, if you start using the integrated graphics of a cpu, the memory will be used a lot more active. Furthermore, I know that in old times AMD had more performance if the latency was lower, where at the same time Intel had more performance if the bandwidth was higher (remember RDRAM? )

 

About the 1600, if you have 1600 CL4 and 2800CL7, then the delay in seconds is indeed equal. However, if you run the memory at 1600 but don't set the 2800 at CL4, which it should be able to run, but leave it at CL7, than the latter is slower.

 

Personally if the PC is for multimedia: gaming, movies I don´t see the need for more than 8GB of memory. If one starts rendering videos\photos then the system actually uses16\24 memory.

 

Yes, for personal usage 8GB is sufficient. Also for rendering most of the time (then only CPU is relevant), only for editing it isn't. When you load a 16MP image it starts to get complicated.

 

 

I'm currently graduation in the field of algorithms, where I work with datasets of 10.000 elements. That isn't that much, however the memory usage for that algorithm is more than quadratic. In my case, I might need to store up to 10.000^2 * log_2(10.000)^2 = 1.600.000.000 data elements worst case. With about 50bytes per data element, that is 80GB. In that case, 24GB isn't even enough. Let's not even consider datasets of 50.000 points, which are about 400 times heavier!

 

I hope I never get that worst case  Also, best case is about 10.000 * 50 bytes.

Edited March 11, 2013 by rolf