RANT: THE CAS LATENCY TIMING DOESN'T MATTER AS MUCH AS YOU THINK IT DOES

hey guys bill zoid here and today we're going to be taking a look at how primary timings especially on ddr5 really don't matter very much now already on ddr4 i've i've complained that people pay entirely too much attention to the cast latency timing which really is there more to look pretty than to actually do anything to your memory performance but you know apparently that message didn't didn't make it through like most of the things i say on my channel um but anyway so here we have a ddr5 setup running 28 37 37 28 1 t command rate like these primary timings are as tight as it gets for ddr5 right now um at 6 400 megabits per second like this this isn't getting any lower uh without like resorting to liquid nitrogen on the memory sticks so for air cooling this is as good as it gets um and the only thing i've not done is i've not touched any of the sub timings all of the sub timings are on auto and yeah we're gonna run some uh latency like just memory performance tests and mostly memory latency tests so we're gonna run ida and we're gonna run uh pi prime uh pi prime is a uh single core cpu benchmark that is so incredibly memory latency sensitive that even right now the world record in it is held by ddr4 it is one of the very few benchmarks out there where ddr4 beats ddr5 um yeah everything else at this point has been at least on hardware ball every benchmark on hardware bot at this point has been taken by ddr5 except pi prime because pi prime loves memory latency and ddr4 just can't beat uh i mean ddr5 can't beat a ddr4 kit doing 4400 cl 12. um so of course we aren't in safe mode so we're going to have to rerun the latency test in ida because it's horribly inconsistent we are supposed to be seeing like 56.7 nanoseconds right with these settings right now which is not bad by ddr5 standards okay we got a 57.9 um you know and our read test is uh just under 100 gigabytes per second our write test is at 89 gigabytes per second our copy test is 90 gigabytes per second um not great as far as i'm concerned in terms of performance the latency test is not being very cooperative right now it is supposed to be able to go under 57 nanoseconds with these settings right now um come on give me okay well but you might be wondering like builds why don't you just take a screenshot of the settings and the thing about screenshots is like people can edit them though i guess if i actually edited my videos i could just edit the footage right and then punch in whatever latency numbers i want um which is one of the reasons like i that's partially one of the reasons that i like to do my videos where like i run the benchmarks on video instead of ahead of time because yeah also this was evidently a bad decision because it's just not gonna spit out the correct result for the next like five attempts or something stupid benchmark i mean we could just call it 58 nanoseconds that's really not that different from 57. so we'll just call that 58 nanoseconds which like it just makes it easier for the second set of time like the cl-40 setup to beat this yeah i'm gonna get better memory latency at cl 40 at the same memory speed because primary timings don't matter anyway next up we're going to run a pi prime and i keep clicking the wrong benchmate icon and pipe prime is going to be very one-sided it's also going to favor the the cl-40 setup in fact everything favors the cl-40 setup because primary timings just don't matter like they well okay that's not true they matter a little bit but they don't matter anywhere near as much as most people think they do so anyway we get nine nine seconds uh 862 milliseconds uh which is actually not the best result for for this setup i i was really expecting to see like a nine nine seconds 800 milliseconds not we've got like 50 more milliseconds than what i want to see um so that's now that's not ideal but oh well it's close enough like the thing is the cl-40 setup is going to have so much better latency that it's going to beat this score by an entire second or almost an entire second so i don't know why i care so much about a couple milliseconds on top of this it's just it really doesn't matter um so anyway now we're going to load up the cl 40 profile which is just going to wipe the floor with the cl28 profile um because again like the cast latency timing is there to look pretty like the trcd timing is actually more important than cast latency is but the thing is trcd on the vast majority of memory chips doesn't go very low and when i say vast majority of memory chips i mean basically everything except samsung b die on ddr4 like i can't even think real well no there was some wacky ddr3 memory chips that would do really really low trcd um but historically speaking like the most important of the primary timings in terms of performance is also that only like primary timing that really doesn't tend to go very low on the vast majority of memory chips ever produced like loose trcd is kind of the define like is just the norm for memory um for basically as long as i've been intimate overclocking it's been a case of like your cast latency goes lower than your trcd and your trcd is more important for performance so anyway now we're going to load up our cl4 actually i'm going to show you that i didn't change any of the of the sub timings because obviously if i punched in uh like if i really wanted to i could punch in sub timings that would send this uh setup straight to hell in terms of performance but as you can see everything is just straight up on auto i've not done anything um and so we're going to load up cl 40 with type sub timings and actually i'm not going to show you the timings because uh like i've these are timings that i've used across many many videos like you can just watch any of my other ddr5 overclock videos you're going to see very similar sub timings to what i'm using here actually you're going to see better sub timings what i'm using here because when you have loose primaries there's a few of the sub timings that don't go as low as they would otherwise go but yeah anyway so here goes the cl-40 overclock that's just gonna straight up b to cl28 at the same memory speed because cast latency just doesn't matter and i get so annoyed when i see comments from people who are like oh well i'm waiting for ddr5 to hit cl28 well you shouldn't be waiting for ddr5 to hit cl28 because it'll still not perform very well because the sub timings will still be trash in fact they might even be worse like yeah like the the only thing the memory like primary memory timings are good for are like determining what memory chips you are likely to get and making purchasing decisions like based on i want to be able to overclock this memory a lot or i want to be able to overclock this memory a little if you get memory like that's why with like ddr4 the focus is on getting 3200cl 141414 ddr4 because that is samsung bdi and it has a ton of overclocking headroom whereas if you get something that's 3216 1818 it could be samsung seady which is complete and utter trash it could also be samsung ddi which is somewhat better than samsung's seed ion the latency test is being as cooperative as ever you can see that the read write and copy tests are way well read didn't really change that much but write and copy are way faster than what they were on the cl28 overclock and oh look we've we've already got a sub 58 nanosecond result um that didn't take very long the best results i've seen from this setup right here is actually 55.1 nanoseconds so we're gonna wait for that but anyway like the you know on the cl28 setup the right test oh my god i just did 54.6 see my point primary timings don't freaking matter if i see one more person complaining about the cast latency on ddr5 being loose i just i'm gonna make another video about how incredibly freaking wrong they are because it's just it's like do you see that i didn't even expect it to hit 55 under 55 nanoseconds i was happy that like this setup was getting like 55 nanoseconds ish whereas i've not been able to get a 55 nanosecond result on the cl28 and i've at this point i've run the ida latency test on the cl28 setup like more times than i run it on the cl40 setup because the cl 40 setup produces better results it just it because it's just better um anyway now if we run pi prime this is just going to get absolutely annihilated um because pi prime you know it like is more consistent and it's also actually like like the thing is any memory benchmark is actually just a cpu test with a memory bias um but um so like it is impossible to make the like a perfect memory benchmark because the thing with any memory benchmark is like that data has to go somewhere and that somewhere is always going to be the cpu cores right so at the end of the day everything is actually just a cpu benchmark with some some level of like memory component but anyway so this is a single core cpu benchmark that just so happens to be extremely memory latency sensitive um and yeah this like this is basically a second faster at cl 40 than at cl28 because this cl 40 setup has way way better sub timings which means the memory latency is actually way lower even though the cast latency sucks because the thing about the cast latency timing is it's only relevant for like a single memory read operation so if you're doing if you do one read command that's when your cast timing gets used if you do many read commands in a row it no longer matters um and there's yeah like if you're chain chaining a bunch of read operations one after the other or even reads and writes and you're just you know making a string of them the cast time the the the cast timing just doesn't do anything um and like even even ida doesn't think like even ida agrees like cl 40 with good sub timings is better than cl28 with auto sub timings like they're not even particularly bad they're just auto because auto timings are freaking terrible um actually well no like that's not true i've seen memory kits that come with xmp timings that are worse than auto um yeah that exists and it tends to happen on like high-end kits in my experience which is really stupid like there's some very expensive ddr4 kits out there which have absolutely atrocious xmp profiles that should just never be used because they're bad they're very bad um like you turn the xmp profile on and your performance actually just goes down um so yeah the problem with ram performance is not the primary timings it's the sub timings um and ddr5 subtimings on auto are really like ddr4 auto sub timings are pretty bad ddr5 some timings on auto are like way worse let's just put it that way i've not actually got gone and like done a direct one to one comparison but like is worse is way worse which also means when you tighten up your sub timings on ddr5 it makes more of a difference to the performance than it does on ddr4 and the primary timings are as irrelevant as they've ever been um arguably they were more relevant on ddr3 and four slightly very slight well the the thing is on ddr4 you have samsung b die and so if you can do like trcd14 right like that that like going from trcd like 20 to 14 that's a big deal um but with ddr5 you're going from trcd like 50 to trcd 37 or like 36 maybe um and your perfor like your performance basically doesn't change because 37 and 40 are pretty much the same number or even 36 and 40 are basically the same number um so anyway yeah uh hopefully this this proves a point the primary timings do not matter so yeah another actually there's like multiple comments that led to this because i remember reading a comment where somebody was complaining that they have like a memory overclock on i think a unifi x where the only thing they did was change the primary timings so of course their memory performance was terrible because they changed the primary timings which don't actually do anything thank you for watching my video like share subscribe leave any comments questions suggestions down in the comments section below leave a if you'd like to support what i do here with actually hardcore overclocking i have a patreon there's a link to that down in the description below there's also the hoc teespring store where you can pick up uh shirts hoodies uh posters you know the usual youtuber merch and i also have a bandcamp um if like you would like to check that out that exists there's a link there uh in the description so yeah that's it for the video thank you for watching and goodbye

Channel: Actually Hardcore Overclocking

Views: 105,806

Keywords: Overclocking, PCbuilding, Buildzoid, AHOC, Actually, Hardcore, Hardware, OC

Id: pgb8N23tsfA

Length: 13min 50sec (830 seconds)

Published: Mon Jul 11 2022