Real World Effects of Memory Timings

[Leckon]

Hey all,

 

tldr; Has anyone tested or know the real world effects of tRCD and tRP timings (the second and third numbers in the x-x-x-x timing listing for RAM) on gaming or benchmarks? Is the effect that they have so minuscule that it is virtually immeasurable? Is it just black magic?

 

Long version:

I'm currently trying to find if anyone has done any tests to find out what the effects that timings have on your system, benchmarks, and gaming. I'm mostly curious about the effects of tRCD and tRP.
To summarize my current knowledge (and to hopefully help anyone else who may come across this post in a similar journey to me):
There are 5 (or 6, depending on if you include a derived calculation) numbers that are associated with RAM.

1. Transfer speed. This is given in megahertz (MHz), and some common speeds are 3200, 3400, 3600, and 4000, though with DDR5 they are getting faster than 7000 MHz.
2. CAS Latency. Also known as CL, tCL, or tCAS. This is the first timing in the x-x-x-x listing that you will see when getting memory. This is the amount of clock cycles that it takes for RAM to be able to "read" a bit.
3. RAS to CAS Delay. Also known as tRCD. Since RAM caches are organized in grids, there is a delay in between accessing a specific column in a row to get a specific bit. This is the amount of clock cycles that it takes to access that bit.
4. Row Precharge Time. Also known as tRP. Since RAM is volatile memory, it needs to have a charge to store data. This is the amount of time in clock cycles that it takes to charge up a new row either initially or after clearing a previous row.
5. Row Active Time. Also known as tRAS. This is the minimum amount of clock cycles that the RAM should take to "write" or "read" data. In quotation marks as writing and reading data is issuing and recalling charge states, but if I use that terminology it gets confusingly close to tRP (because it is).
6. The previously mentioned 6th number that you will read about (mostly on PCPartPicker), called First Word Latency, is derived from the CAS Latency and the transfer speed using this formula: CAS Latency/(transfer speed/2000). This gives you, in nanoseconds, how long it takes for the RAM to return a bit. (Side note, this is apparently technically not correct, but this is what PCPartPicker uses as one of the primary categories when sorting RAM, see last link in sources to someone who explains it better than me.)

If anyone has any simpler explanations or analogies to help me or anyone else coming across this thread please share, greatly appreciated!

 

Sources for my maybe incorrectly summarized knowledge after a night of rabbit holing:
https://gamersnexus.net/guides/3333-memory-timings-defined-cas-latency-trcd-trp-tras

(Video link for the same article:)

https://ro.pcpartpicker.com/forums/topic/364651-dram-first-word-latency

[RONOTHAN##]

1 minute ago, Leckon said:

I'm mostly curious about the effects of tRCD and tRP.

tRCD does a bit more than tCL on average, tRP does next to nothing, and tRAS does even less. You usually need a pretty big change in tRCD and/or tCL in order to actually notice it, 20 vs. 22 or 36 vs. 40 is going to be effectively the same performance. 

 

2 minutes ago, Leckon said:

CAS Latency. Also known as CL, tCL, or tCAS. This is the first timing in the x-x-x-x listing that you will see when getting memory. This is the amount of clock cycles that it takes for RAM to be able to "read" a bit.

So I will critique this definition a bit, as while it's technically accurate, it doesn't take into account how the memory controller uses CAS latency. The memory controller can cue up memory operations, and so that CAS latency value is only actually applied once for a set of read operations. If you're doing a bunch of random operations in the memory, it can be harder to cue up memory operations so you can actually mask the effects of tCL, but it's still one of the easier timings in the memory system to mask. 

 

If you do want to see a bit more of these, Buildzoid did a pretty decent series where he went through a few of the more major memory timings and actually showed how they would function in real time if you were a memory controller trying to read data from the memory and thus how it would affect performance. 

 

[Somerandomtechyboi]

27 minutes ago, RONOTHAN## said:

You usually need a pretty big change in tRCD and/or tCL in order to actually notice it, 20 vs. 22 or 36 vs. 40 is going to be effectively the same performance. 

i know the primaries dont give that much of a performance bump but damn thats pretty depressing considering cl loves to gobble voltage though i assume benchmarks like spi will probably still show a slight diff

 

im curious if higher freq will show a bigger difference between loose and tight primaries, maybe it isnt noticable at 3800 but maybe its noticable at 4800?

[RONOTHAN##]

12 minutes ago, Somerandomtechyboi said:

im curious if higher freq will show a bigger difference between loose and tight primaries, maybe it isnt noticable at 3800 but maybe its noticable at 4800?

Not really in my experience, higher frequencies scale timings about the same, if anything slightly worse since you have more clock cycles to mask bad timings. 

 

16 minutes ago, Somerandomtechyboi said:

i know the primaries dont give that much of a performance bump but damn thats pretty depressing considering cl loves to gobble voltage though i assume benchmarks like spi will probably still show a slight diff

I will mention if you change both of them, the difference can start to actually become slightly noticeable, but it's still not that large.

[Alex Atkin UK]

2 hours ago, Leckon said:

Transfer speed. This is given in megahertz (MHz), and some common speeds are 3200, 3400, 3600, and 4000, though with DDR5 they are getting faster than 7000 MHz.

Actually its not, memory speed is in Mega Transfers but for simplicity almost everyone calls it Mhz.

 

As memory is DDR (double data rate) then the REAL Mhz speed is actually half the MT speed, as two transfers can be done per clock cycle.  This often catches people out as tools like CPU-Z may display the real Mhz.

[Ebony Falcon]

Real world outside of benchmarks and if you turn the fps counter absolutely 0 

 

as cool as it is to benchmark if someone turned my pc back to xmp I’d have no idea about it 

125 to 150 in the lows was no perceivable difference although cool to see in the benchmark