Performance impacts by CAS Latency

[Adi Avraham]

I was looking at RAM and found a kit of 2x4 DDR4 sticks running at 4266MHz but at CL 19. Would DDR4-3200 MHz sticks with CL16 perform better in-game? Also, what applications will be most affected by high CAS Latency? Thanks in advance for any answers you can offer.

 

[The_Staplergun]

So, the changing the cas latency will make a difference in nanoseconds per access request. This, overthe span of the clocks per second, will probably amount to not even a second at ddr4 rates.

Your biggest helper will be bandwidth. Don't just loosen the cas timing out like crazy, as the effect can be a little bit more apparent at higher frequencies. It's still well within the span of a second or two if something crazy happens.

Bandwidth allows for more calculations per second. The CL is how long in system clocks it has to wait for the data to be available from the memory bank to read, after sending up its request. Because the way it works though, you can have requests going up while others are and while other requests are coming back for being available is a simple way to put it. So really cas latency in effect would really only effect right at the start of loading. Once the ram knows where it's being accessed it can read much faster sequentially because it knows what it needs next so it can predict.

CL also really effects random reading and writing, which is generally from initial loading if I understand correctly.

Also, for ddr4 ram, still negligible.

Bandwidth is your friend. Higher frequencies for life.

You would see more of a perceptible difference between those two frequencies than you would dropping the cas latency by a point or two at 3200mhz. Albeit, the only times you'd see enough of a difference are when the ram is attempting to load very large files, in the multiples of gigs at once. Even then it's not by a whole lot realistically. Maybe up to a few seconds instead of 1 or 2.

 

[steffeeh]

As stated above CAS latency is latency calculated by RAM clock cycles - not real time.

To get the real time latency:
1/((RAMspeed/2)/1000) x CL = True Latency in nanoseconds

First we need to calculate the duration of a clock cycle. Take the RAM bandwidth and cut it in half, as the actual RAM clockspeed is half of the bandwidth (or speed specified by the XMP or manual speed setting).
Then divide it by 1000 to convert it to GHz -> to make things easier as GHz represents X billion times per second, and a nanosecond is one billionth of a second.
Then when you're done, divide 1 by the value you have to get the duration of the clock cycle. Then multiply it with number of cycles of latency, and you have the entire duration of latency in nanoseconds.

1/((4266/2)/1000) x 19 = 8.91 ns
1/((3200/2)/1000) x 16 = 10 ns

Not only is the 4266MHz memory faster in speed, but also shorter in real time latency, and therefore better in every aspect compared to the 3200MHz stick, given that it will actually work in your system (taking into account; mobo memory support, chipset compability, compability with CPU clockspeed, etc).

Anyway, as for CAS latency vs speed - speed is mostly preferred for heavier applications, as it cuts down rendering time (even if it's only a fragment less), as well as high framerate gaming such as 144Hz gaming or even 240Hz gaming (given that you have proper specs on CPU, GPU, etc), as it helps increase the minimum FPS floor and stabilize the average FPS. However this is still a bit controversial, and I've heard people say that above 3000MHz you get diminishing returns as you go higher.
You only want 3200MHz over 3000MHz as the comability for the latter can be somewhat dodgy sometimes.
As for CAS latency however, I'm not sure, but I've experimented with it myself pushing it quite hard, and both me and a friend trying it out as well felt the overall impression of manuevering the PC felt much snappier and straight to the point. However I can't guarantee there was an actual improvement.

 

[The_Staplergun]

Yeah it could be seen as the placebo effect unfortunately. There's a great chart on Wikipedia that shows tons of times in ns for different frequencies and ram types.
https://en.m.wikipedia.org/wiki/CAS_latency#

 

[sdancer75]

As stated above CAS latency is latency calculated by RAM clock cycles - not real time.

To get the real time latency:
1/((RAMspeed/2)/1000) x CL = True Latency in nanoseconds

First we need to calculate the duration of a clock cycle. Take the RAM bandwidth and cut it in half, as the actual RAM clockspeed is half of the bandwidth (or speed specified by the XMP or manual speed setting).
Then divide it by 1000 to convert it to GHz -> to make things easier as GHz represents X billion times per second, and a nanosecond is one billionth of a second.
Then when you're done, divide 1 by the value you have to get the duration of the clock cycle. Then multiply it with number of cycles of latency, and you have the entire duration of latency in nanoseconds.

1/((4266/2)/1000) x 19 = 8.91 ns
1/((3200/2)/1000) x 16 = 10 ns

Not only is the 4266MHz memory faster in speed, but also shorter in real time latency, and therefore better in every aspect compared to the 3200MHz stick, given that it will actually work in your system (taking into account; mobo memory support, chipset compability, compability with CPU clockspeed, etc).

Anyway, as for CAS latency vs speed - speed is mostly preferred for heavier applications, as it cuts down rendering time (even if it's only a fragment less), as well as high framerate gaming such as 144Hz gaming or even 240Hz gaming (given that you have proper specs on CPU, GPU, etc), as it helps increase the minimum FPS floor and stabilize the average FPS. However this is still a bit controversial, and I've heard people say that above 3000MHz you get diminishing returns as you go higher.
You only want 3200MHz over 3000MHz as the comability for the latter can be somewhat dodgy sometimes.
As for CAS latency however, I'm not sure, but I've experimented with it myself pushing it quite hard, and both me and a friend trying it out as well felt the overall impression of manuevering the PC felt much snappier and straight to the point. However I can't guarantee there was an actual improvement.

I think that the formula

1/((RAMspeed/2)/1000) x CL = True Latency in nanoseconds

is correct only in the DDR. In the DDR2 the formula should be

1/((RAMspeed/4)/1000) x CL = True Latency in nanoseconds

in DDR3

1/((RAMspeed/8)/1000) x CL = True Latency in nanoseconds

and in DDR4

1/((RAMspeed/16)/1000) x CL = True Latency in nanoseconds