It would be interesting to know which manufacturers have a history of having crap XMP settings because that would definitely be a strike against them in my book.
In my experience, ADATA and almost every other manufacturer that's smaller than ADATA. Disclaimer: not all kit from ADATA and/or smaller manufacturer is crap, I mean some of their kit CAN BE BAD.
I mean what he said in the video was that stability could be a trade off. So depending on how this works with my tiny understanding of it I would actually gladly sacrifice one step down in sub CL for stability for my gaming rig. I mean lets say it drops performance by 5% then with all of the other components doing their thing maybe in the end we are talking 0.1% lower fps for extra stability. As an overclocking enthusiast I understand him thinking they are garbage but there is definitely a place for them.
@@HOTTESTHERE Unfortunately it's not simple math like 5% lower on ram and other component makes up for it.... In reality it's CPU waits for data in L1 cache, if it's not there then L2, then L3, then RAM, effectively CPU need to wait until it get what it want before processing, this happens in nano second level so you won't be able to tell. while CPU stalling waiting for DATA, GPU will also stall waiting for CPU's instruction of what to render next.... Depending on the game, if the game was coded with good optimization, where the CPU are able to predict what data is supposed to get pre-loaded to L1/L2/L3 cache ahead of processing, RAM does not matter at all, though games right now usually have way too many players/props, and the CPU is unlikely able to predict what to shuffle into its cache, hence memory performance is a HUGE factor as of right now. You will notice this especially noticeable in battle royal games. In RAM the data are stored in bank, bank are grouped to form a Rank, then Dimm.... so when some memory kit come with "crap XMP settings" it usually means the default tRRD/tFAW is bad. That represent how long of a delay to get info from banks.
@@OMGJL I've made the mistake of buying ADATA RAM in the past and I think once was enough for me to realize that it's a crap brand. Any enthusiast should know really..
It depends really on the quality of the chips that are binned. Small manufacturers can have great modules. Large manufacturers can have crap. It all depends on the chips from the lot batch.
I think it would help a lot if motherboard manufacturers just displayed the resulting latency in nanoseconds next to the timings in the BIOS. Not only would that point out the difference to people who dont understand, it would also be helpful so you dont have to calculate them manually to see if something is an improvement or not.
They do that's the cl number, that's what he explains in this video that's the general nano second timing of the ram But if you over or under clock the cpu besides the ram, the cl times could change. As usually the cpu is the ram controller.
@@zagan1 I've never seen a bios that does so. They all just display it in clock cycles. And no, the memory controller/cpu has nothing to do with that at all. I'm talking about just the individual latencies, not some resulting latency that would be more complex to calculate (even though that would still not be hard).
@@llortaton2834 untrue. No one will upgrade but people new to the space will be basically forced to get DDR5 since chipsets are moving to it. DDR4 supported chipsets are still on the shelves but not for long. Had to buy a used Z97 only a couple years after DDR4 came out since I couldn’t find a new DDR3 Z97 that didn’t have a jacked up price. New users won’t want to buy used.
The fact that cas latency wasn't indicated in ns always made me aware that there was more to it, but nobody (at least big) talked about it. Thank you very much
As a matter of fact back when memory chips were SIMM instead of DIMM, they always reported latency in nanoseconds in the sales catalogs. Back in those days there did not exist any bios settings to set your own memory latencies. In any case low memory latency really only becomes important at high framerates anyway and 60fps has always been plenty for me so i never spend more just for lower latency at a given clockspeed. but i do spend more for higher clockspeed ram. And sometimes i spend less for higher clockspeed ram. ☺️
@@Sidicas good point. In fact for the past 4 years I consider everything below 144fps to be low fps. This understanding of cas latency made me explore what previously I thought wasn't possible with my current ram config. Atm I'm still validating for stability, but already gained 7-8% of average frame rate (161 to 174 on ow, 5800x with 1660 ti). Yes I'll soon upgrade the gpu, for sure it's my bottleneck atm.
@@walls89 I guess the question becomes, where do we draw the line between doing it for envy(it's not jealousy), and doing it for accuracy? And if BZ was on the off chance wrong in an instance like this, does that suddenly shift it over the line? No, that doesn't sound right. So then what would that line be? I don't really know. I guess maybe it's a; you know when you see it type of things. I tend to think of BZ doing it because he wants to add clarity in the world when the opportunity arises. That's how I like to look at it, and try to do at at times.
No way buddy, you should correct misconceptions wheneveryou can. I've been a PC/Net Tech for roughyl 20 years. I've been fiddling with PC's since the late '80s, and I was unaware of exactly how to figure out the *exact* difference in speeds. I was way, way off btw, using a _common sense_ method which was apparently way off and similar to the way _"The Good Ole Gamer"_ thought it worked as he talks about it all the time. He's a much smaller channel, but when you see this stuff, and you *know* they're making an oops, you potentially have the chance to teach them, and the hundreds if not thousands of people who actually care about these things and *NEED* to know in many cases because it makes us better at our jobs. I honestly thought pc3600 cl16 was faster than PC 5200 cl40. Making that mistake thinking I'm saving a consumer money, or saving on my own rigs, is a big deal. Lying to a customer unintentionally is a really big deal, so please, continue to correct mistakes or misconceptions when you see them. I'm 99.99% sure Linus actually appreciates it. You aren't presenting it out of malice to belittle people. You present it in a way that clearly states you're doing it so people know and get the best performance they can, plain and simple. You're too laid back cool to get pissy with anyway. Thanks for teaching me something I was unintentionally ignorant for...well years apparently. Good show.
Also, you shouldn't put "PC" in front of RAM speeds unless you're specifically adding the generation and bandwidth to that old unit of classification. PC 5200 makes it sound like you're talking about DDR2 667 (PC2-5200). If you were going to talk about DDR5 5200 in terms of "PC", you would say PC5-41600. It's a number system which is rarely used today but could confuse people who were used to seeing RAM advertised by PC number back in the day.
@@tired9494 Yeah memory generations shouldn't be directly compared this way. DDR3 has inherently lower latency than DDR4 and DDR5 is even worse but that's not the whole story. My best DDR3 rig has a CAS latency of ~7ns but it isn't going to outperform DDR4-3600 CL16 @ ~9ns unless I double the output by running on a quad channel platform.
I think the confusion happened because 3600 Cas 18 is the same as 3200 Cas 16 and because those were two very common pairings people treated it like the golden rule of ram.
It's actually a much easier problem than even Buildzoid is making out, but we get lazy with our shorthand. Because it's just two factors multiplied together, you can just look at the ratios to determine which is greater. For example, 18 to 16 is 9:8 (factor of 2) and 36 to 32 is also 9:8 (factor of 4). So we know they are the same. Whereas 42 to 38 is 21::19 (factor of 2) and 40 and 38 is 20::19 (factor of 2). So, with some simple math, we can already tell that the difference in clock speed is greater in ratio than CAS latency and thereby the difference in better vs worse.
TBH, the way I always did it was just freq/cas. I found it more annoying to do (1/(actual clock))*cas. Doesn't give me real value in ns but allows me to quickly make comparisons.
At least if you do cas/clock you'll get the correct numbers.. the unit might be off depending on how you type it tho. E.g. 40/2.6 = 15.38ns 40/26 = 1.538 40/2600 = 0.01538 microseconds
That 1/(actual clock)*cas does give right value, just in ms. So, if you first divide the actual clock with 1000, you get the answer in ns. You could also use calculators [1/x] button, i.e. 1800 [1/x] is same as 1 [ / ] 1800 [ = ]. For example, *DDR4-3600* (15-...) --> CAS Latency is 15 clock cycles, I/O bus clock 1800 MHz: 1 / 1800 * 15 = 0.00833 ms (button presses: 10) 1800 [1/x] * 15 = 0.00833 ms (button presses: 9) 1 / 1.8 * 15 = 8.33 ns (button presses: 9) 1800 / 15 = 120 (button presses: 8) 1.8 [1/x] * 15 = 8.33 ns (button presses: 8) 1.8 [ R ] * 15 = 8.33 ns (button presses: 8) Windows Calculator: clicking *[1/x]* with a mouse is same as pressing *[ R ]* on the keyboard.
That's was a really good explanation of timings for some reason I just cannot comprehend it when reading through a forum or a wiki post, videos like this make it easier to understand for whatever reason
Am I missing something? Your video just showed exactly what he said. What he said was going eg 4800 CL38 to 5200 CL40 will get your roughly the same latency. He didn’t say timings are an independent measure of time. He simply worded “a higher CAS value” poorly but the point was clear.
Linus said that one kit had lower latency and the other had higher bandwidth, so it did not matter much what was chosen. But the calculations say that the corsair kit has both lower latency and higher bandwidth, even also a lower price, so the Corsair kit is better in every way.
The comments are absolutely trashing Linus lmfao. I love his content and he makes entertaining videos. Doesn't mean I have to listen and believe everything he says. People make mistakes it happens. Whys everyone on the internet have to pick sides or trash someone who said something wrong or made a mistake as if everyone doesn't make mistakes. That's why we have other channels like this for in depth specific information and alternative content. Everyone has there strong suits. Everyone takes everything way to seriously today. Literally saw comments saying they're done watching Linus after this like wtf lmfao.
@pabz yea confuses the hell out of me! Especially since he proved Linus right! He said that anything under 10% is not perceivable by a human and then says the corsair ram is faster in every way, which is true, by 8% in bandwidth and 3% in latency. I'd qualify that as "quite similar" which is what Linus said.
@@christopherjames9843 So your solution is to be perfect? What planet are you from. Yes what you're saying is technically true but that's life. People with platforms are still people like you and me and they're going to make mistakes. No reason to jump ship and grab your torches and pitchforks to riot. Relax.
@@PabzRoz huge false equivalence fallacy there bud. He's not claiming he has to be "perfect". This was a situation Linus could've easily just fact checked himself before hand, a easy basic thing someone of his size should be doing. That's not being "perfect" that's doing the bare minimum.
@@krisv1038 Fact checked himself before hand during a livestream where the topic randomly popped up... Right makes sense. He should of been able to tell the future and known memory timings and latencies were going to be discussed and used his calculator to do the math so he had an answer ready. Dude do you even hear yourself? It was a mistake bro. Get a grip. "RU-vidr says something wrong omg let's all pick up our torches and pitchforks"🤦♂
great piece ..seems there needs to be a bunch of videos and articles about ram and how it functions ..loved how you kept it on topic and didnt bother roasting linus too much its easy to drill into someone but to educate us all is much more valuable ..ive been building for decades and still dont properly know how ram does its thing
I get where he's coming from, since with DDR4, 3200CL16 has the same effective latency as 3600CL18 (all else being equal). But that only works with specific kit values, so you can't apply the same generic logic to all speeds without actually calculating effective latency, like he failed to do.
EDIT _ What about the exact speed difference between a 3600 CL 16-19-19-19-36 kit, compared to 3600 CL 18 ...something? Or, 3600 CL 16-19-19-19-36 compared to 3600 CL 14-14-14-28 ??
@@Dennzer1 Since the speed is the same for all those examples, the lowest latency kit IS the one with the lowest effective latency as well. The only time effective latency matters is when you're comparing two different speeds. When they're the same speed, you just compare the absolute latency.
@@bjn714 I'm sorry... I was just trying to ask which was faster for gaming at 1440p, HOW MUCH faster is a 3600 CL 14-14-14-28 kit compared to a 3600 CL 16-19-19-36?? And also how much faster is 3600 CL 16-19-19-36 compared to 3600 CL 18 ( and for the rest... tighter and whatever the most common 3600 CL18 sub timings are, and especially NOT those awful 18-22 etc. kits... ) ? As BZ himself says, CL does not mean Latency even though it is in the name. I was just trying to ask which was faster for gaming at 1440p, but I have to apologize for wasting your time with a poorly worded question...
@@Dennzer1 Again if the speed is equal, the lowest CL will have the lowest effective latency. So the 14-14-14-28 b-die kit would be the winner, plus it would have plenty of room for tightening subtimings to extract further performance beyond what the XMP profile will set/provide.
Thank you Buildzoid. These last few RAM videos have been very enlightening. A lot of things I thought I understood were not quite right, and I'm glad to know what I was getting wrong. If there's one thing I like, it's feeling less stupid at the end of a video. Awesome stuff!
I just want to say thank you. I can't tell how many times I have watched videos and seen things that have been stated wrong. Then when you read the comment or go on forums you see people post replies using big name people and think just because they said it was this or that than it has to be correct. I have called out a few in the past an then I get flamed in the comments for saying something. So, it's nice to see someone else calling people out.
I also use this for example 40/5200 x 2000 = 15.3846154 As long as the sub timings are proportionate So the corsair kit has a better first word latency stock not accounting the die
honestly i dont think a lot of techtubers would even know about calculating latency that way. it would be nice if you guys can emphasize about this more.
Linus said, that the Kingston kit (4800 CL38) is slightly slower compared to the Corsair kit (5200 CL40): 0:55 "... and Kingston which has a slightly slower kit...". So this statement of him was not wrong, no need to "STOP". But I think you didn't understand what he meant with his "binned kits" statement: As you calculated, 4800 CL38 has a very similar latency in nano seconds compared to the 5200 kit. And that is exactly what he meant: If you go a "step" up with the "MHz" (frequency), e. g. from 4800 "MHz" to 5200 "MHz" and increase at the same time the CAS latency from 38 to 40, you will get similar performance. In other words: If a kit with higher frequency but also higher CAS is cheaper, you can grab this kit *but* only the difference is only "one step" in the same direction. You could prefer a 5200 CL40 kit instead of choosing a 4800 CL38 kit, no matter if you really know the exact latency *timings* at all. They only question is: How the "stops" (like in photography) are defined: What are those steps in general: 4800 -> 5200? Or is there something else in between you need to consider? E. g. Is a 5000 CAS 39 faster as a 5200 CAS 40 (in case odd CAS values are usable at all of course)? A buyer needs to know: 4800 "MHz" is in truth half of the value -> 2400 MHz only, so the "steps/stops" are always to be considered in 200 "MHz" steps: 4800 -> 5000 -> 5200 and so on. But is the CAS stepping also always 2 or is it 1 and therefore the usual "stop" is a diffrence of 200 "MHz" and 1 CAS?
He didn't say the cl38 was better. He said lower clocked memory with lower cas timings can be as good or better than faster clocked memory with higher timings. He's not wrong. Hardware Unboxed has tested this showing low timing 3200mhz is as good as high timing 4800mhz memory in most instances.
You're missing the point. Higher timings and higher clocks can in some circumstances have lower absolute latency which pokes a hole in Linus' universal statement latency = clocks / timings the simplistic misconception that clocks = bandwidth and timings = latency that I too was once guilty of, is wrong
@@greebj I thought he was just saying that sometimes lower timings can make up the difference of slower clock in some circumstances, which is true. Cl14 3200mhz might have lower latency than cl34 4800mhz for example.
What makes me sad is how difficult it is to custom tune timings; There are so many different sub-timings, I feel like I need a PHD or Masters to understand how they all work together. The other side of this coin is marketing. CAS is used to differentiate different memory kits from each other, so of course everyone is going to think they're still relevant.
Most of the time for regular users you just push the voltage to the recommended limit, then tighten the mhz, then tighten the cas latency and be gone with it, because the chances of getting a crash in the future outweigh the negligible improvements you'll get
Sure there's a lot to take in compared to other hardware fine tuning, but still - if you look at various forums, you'll quickly notice how people first most are ignorant. Let's take Ryzen PBO2 curve optimizer for undervolting: dude slams -30 on all cores, runs one pass of Cinebench and calls it stable 🤡 - and it's relatively basic process compared to RAM overclocking when you have not just bunch of timings but also other parameters. So if you don't have time digging thru it - you simply buy kit with optimal XMP for given CPU - so for example for Zen3 a 3600/CL16 is optimal and still not too expensive. Also would be worth noting that generally min-maxing is utter waste of time (unless you do it for competitive benchmarking) - so even if you manually overclocking - slam 3600/CL16 with decent timings and call it a day and it will be super quick and easy to stabilize (likely 1st try). But if you push it say 3800 (and start playing in territory CPU IMC capability limits) and pushing super tight timings (and when I general say timings, I mean all of them) - you spend days on 100% stabilizing it and it still there is uncertainty of it spewing errors in certain workload (might be very specific game) and for what? like 1-2% performance gains which is literally on margin of error when you're benchmarking something - so congratulations on wasting days of your time for nothing. Most people simply do not understand diminishing returns - with any overclocking there is a point after which gains are negligible, but time need to tune increases exponentially. Same with PBO curve optimizer - if you tune the curve in increments of 5 offset points - it will be quite fast process and you'll take 90%+ gains that are on the table. But if you start finetuning to a single offset point - you immediately multiply work by like 5 times, not to mention it's also harder to test for stability, because if we assume -17 offset is 100% stable, you slam -18 (because -19 wasn't stable) and you may not catch instability with synthetic testing as it's so much almost stable. However if you operate in increments of 5. So you slam -20, that is obviously not stable, you drop it to -15 and bingo, 100% stable, and difference between -15 and -17 will be absolutely imperceivable neither in performance nor temperatures. So in other words, stability certainty is much higher with bigger increments, whilst in first case you need to go -20, to -19, to -18, to -17 with ever increasing testing times on each step as the closer to 100% stability you are the harder is to catch errors. tl;dr - don't overdo it, because it's waste of time.
@@ActuallyHardcoreOverclocking People are willing to stress test their mhz or CL overnight but i've seen them skimping on the other timings after they've failed several times and end up stressing for half an hour or even less
Oh i remember this part. Anthony probably knows better and could've corrected him but as long as Linus doesn't recommend terrible overpriced memory kits or regurgitates manufacturers's BS marketing i'm happy. I still remember his disastrous 3080ti video and what was basically giving Nvidia a pass to rip people off.
@@JimJamMS eh, I thought it was just lazy and out of touch. I've always seen Linus as a nice arrogant asshole. I've worked with a few people like them. They usually admit they are wrong eventually and to a degree and Linus does that most of the time. The fact he told Nvidia to go fuck themselves when they blacklisted hardware unboxed is good evidence he isn't just bought and paid but his enterprise is a content making machine and like all mass produced shit, defects creep in now and then
linus points on his 3080 Ti review on wanshow were valid, its just people can’t get along anymore without declaring war on everyone that has a different opinions
Linus' 3080 Ti video was fine. It was thought-out, decently explained and it was just his opinion. Getting CL timings wrong is worse because it's a statement of fact and it's demonstrably false. He probably should have known too because he has presented multiple videos in the past explaining RAM speeds and performance. Some of those were of tests done by Anthony, who definitely should have known and corrected Linus.
Good job. Explain the ram timing chart which shows the segmentation of the clock cycle including rise/fall times and how they add up to be important to not omit when/if you calculate timings... likely some spudding engineers in the audience. And everyone would probably be interested in DDR vs old non-ddr in how rise/fall time logic transitions got decoupled as a pair for the logic provided (doubling of bits for 1 logic waveform... not 1 clock cycle). It has to work this way for the timing values mfg's write to the non-volatile ram in db lookup... not saying they do it well. But has advantage for mfg's as they can create 1 stick and just write the timings they want after binning... if they bin. Less costly work thus more return. or Just RAM it in their ears ; ) Have a great one BZ.
You're pointing out something very essential. But I also think this is where this statement comes from. For random access the latency should be the dominant factor. If you keep latency constant (and thus have similar performance for random access patterns), frequency and latency timing step up in tandem. But sure if you really want to compare it, you need to calculate.
@@reviewforthetube6485 “I resemble that remark” is a play on “I resent” where the person playfully self deprecates by acknowledging that the criticism applies to themselves. But some people are just a stick in the mud and assume they are better than others while being confidently incorrect.
When I look at latency, I've started observing the "First Word Latency" on pcpartpicker. How correct is this measurement on determining the actual latency of the kit?
its neither correct nor wrong.. 1. There's more timings other than first world latency. like tRCD tRP tRAS tRC tRRD tFAW etc etc etc 2. you won't know which's better by just compare first world latency, for example one kit with 4000c14 with tFAW set at 40, vs one kit with 4000c16 with tFAW set at 16, obviously the ladder is the better kit, contracdict to the first world latency figure. 3.DDR5 naturally have worst frequency to latency ratio, but DDR5 have double the bank group hence potentially better real world performance (depends), you can't compare the nanosecond between DDR4 to DDR5..... 4. same as reason 3, you have 2 identical kit in terms of frequency/latency, but one was x8 and one was x16, using your nanosecond method, they will appears to be the same. (hint: x16 is DOG CRAP) But it's still a decent tool to determine which memory have good ICs... like with DDR4, if you select memories with first world latency below 8.5ns, you will almost only end up with at least okay-ish quality Samsung B-Die, basically everything execpt Crucial Ballistix Max (Micron ICs)
first word latency is derived the amount of time it takes [CAS latency number] of clock cycles to happen under the raw ram frequency example: 3200cl16 RAM - 16 cycles at 1600mhz (DDR = double data rate) takes 10ns to occur (at 1600mhz, each clock cycle is 1/1,600,000,000 of a second = 0.625ns) It's a good rough idea of how decent the latency is on the kit. Generally speaking lower is better. But if you really really care about getting the lowest latency possible you would be looking at the other timings as well, or trying to figure out what ICs the RAM has so you can get a good overclock on it yourself.
I think it really comes down to the same thing being worded wrongly. What they meant (i assume) is that 1 ram stick with more Mhz but "Higher" (longer) CL has a similar actual latency in ns compared to a kit with a lower Mhz but shorter CL. But agreed, in the end it all comes down to the actual time latency in ns which is often never communicated at all.
Linus is a salesman and a showman. He is great at talking to the camera. What he does need to do is pass technical advice to his staff who have experience. When he goes off script it all falls apart especially with regards to technical advice.
i currently have 4 trident z royal sticks DDR4 3200 cl 16 . (got them for the look) i noticed one kit has samsung chips and the other hynix. i'm not going to act like i know a lot about ram . what would your recommendation be?
I can finally say I understood one of your videos, mainly because this is one of the very few things I already knew about RAM and the formula I always use when making a purchase decision. P.S. I managed to get my Balistix 128GB RAM to run at 3600MT/s 14-17-17-36, seemingly stable with tight sub-timings. DRAM voltage is set to 1.5 without touching other voltages, my 13900K is limited to 253W which is how I want to run it. All thanks to the knowledge I've gathered from watching your channel in the past few days.
It would be nice to have a big roundup on AM4 socket. All chipsets we have now for years. Is it better to go with B350 for Zen3 CPUs just because some MBs have better build quality or X570 particular boards. That is the question i often ask myself. Thanks for your work, your videos really helpful.
We're talking about a 3% difference in latency between those kits. Is that really worth mentioning? He didn't say CAS timing determines latency either, what he said was that a lower bandwidth RAM kit with proportionally lower CAS timing is going to give you pretty much the same latency, which is true.
Yeah, and with ever increase bandwith and frequency, common latencies doesn't matter as much. Subtimings though. They affect SO much on performance. And especially so on DDR5.
Didn't watch your videos in ages, but I saw your title and new it was gonna be good. Ram needs time to understand. Recently I've seen RU-vidrs use the term megatransfers more and more recently(past couple years). But the understanding usually stops there.
Ian Cutress from AnandTech called out Linus on use of MHz instead of MT/s for RAM. Linus made a snooty "I don't care" reply video, but it's good to see he now fixed his use of that terminology at least. Maybe he'll bless you with a similarly snooty reply! P.S. I work in chip design, and it's common for us to quote 'latency' in clock cycles. It's implied that you mean number of clock cycles at the maximum allowed clock frequency, hence equivalent to time.
And Linus it's right, 99% of people don't care about this. The other 1% watch this channel and tech tech potato (and live in a basement, but that's beside the point).
@@Eduardo-eg9bfLet me correct you: 99% of people don't know what RAM is. Of the remaining 1%, 99% of them grasp the basic concept of memory speed but don't know what it is really about. The remaining 1% of the 1% of people watch this channel and tech potato, and get mildly annoyed by people getting it wrong.
Oof. 2022 and people dont know that cas latencies scale with ram clock speed. Hugher cas at higher clockspeed can be the same latency as a lower cas at lower clockspeed. Geez i remember learning about that back in 2000 or so when the first guides for overclocking those kinds of ram settings were written.
btw when i run 3D Mark my screen goes black and never comes back, until i restart. and my time/day change so i have to update them aswell. is this motherboard fault or psu? i think without systeminfo it doesn't happen
Hi. Was wondering if you knew this! I'm looking at 2 x 32gb 5600m/t DDR5 at cas28. How does this scale if I add another kit - 128gb. Is it advantageous to go the fast cas route.. How does it scale? Thanks Malc.
what ddr5 kit should i buy to replace my ddr5 5200 cl40 corsair dominator?? i cant add voltage to it at all or tweak it at all its 75NS with 89k read, i have a z690 unify X. Im contemplating ordering Gskill 6400mhz cl32 but i wanna hear your opinion. i dont care about money i want sub 60NS. This ram is killing my 12900k and 3090ti
Nice to point it out, I did know about CL is about cycle, but I didn't calculate that, so I would not think about that this kit could have lower latency (as time).
The website I use for comparing memory kits actually does the calculation and shows it as a "True Latency"-variable. It's not that hard to measure yourself. He is right about the "speed bin" thing though. Having a lower CL on a lower clockrate can sometimes be faster than a higher CL on a higher clockrate
Never been good with Numb3rs but I have the g.skill royal elite 3600 16gb x4 cl14 14 14 kit for my TR3960x on zenith alpha board and it may be placebo but it seems to run better than the 3600 cl 16 19 19 kit I had before. So 1÷1.8×14 = 7.77. So 7.7CL?
Here's a simple shorthand. Look at the ratio, rather than the absolute values. For example, two popular kits of DDR4, 3200 mt/s CAS 16 and 3600 CAS 18. 18 to 16 in CAS latency is 9:8 (factor of 2) and 36 to 32 (in hundreds of transfer per second) is also 9:8 (factor of 4). So we know they are the same. Whereas 42 to 38 is 21::19 (factor of 2) and 40 and 38 is 20::19 (factor of 2). So, with some simple math, we can already tell that the difference in clock speed is greater in ratio than CAS latency and thereby the difference in better vs worse.
You should make a crash course vid on how ram works from beginning to end. Unless you already have and I've missed it. I'm really interested in learning. I have 2×8Gb 3200 16 team group xtreem ddr4. I would love to know how to set the timings but I don't know what any of the values and multipliers do.
Anyone have secondary timing tips for the g.skill 3600/cl16 2x16gb b-die kit on 11th gen? I’m running it at 3600 (couldnt do higher gear 1 than this) with 14-14-14-28 2T (couldnt boot with 1T) and 250 trfc / 1.49 dram volt
I'm building a pc soon with 12700k 3070ti. Which should I get 32gb ddr4 3600mhz cl16 or something faster than or equal to 5200mhz ddr5 ~cl40. I wonder if taking ddr4 now will be bad in long run as I'm planning for no upgrades for 4 years but budget is tight
If you can get dirt cheap ddr4, you’d might as well imo. 13th gen is the last LGA 1700 socket generation, and it supports ddr4 as well. If you’re buying high end DDR4 (ie specifically aiming for B Die kits), you’d might as well spring the little bit of extra cash on a ddr5 kit imo
@@zuckdaddy1596 Ddr5 is exact double the price here for 6000mhz vs 3600mhz. And 5200mhz kits are only slightly less than two times. Honestly I don't even know what B die kits are, I heard they oc well, I was thinking to buy the cheapest 32gb ddr4 3600mhz cl16 kit I can find and then trying to oc it to 3800mhz on same or slightly better timings, nothing fancy
if Regional pricing for DDR5 is way, way higher than nice DDR4 Kits, that makes it an easy decision, doesn't it. come several Years there will be way better DDR5 available so if you bought a DDR5 Kit now, since you're looking at AHOC in the first place, you'd..... probably be Buying another Kit then anyways. so in the end you'll be Buying something else in the future no matter what you pick now. for what it's worth adjusting nice DDR4 is way less hassle. and if you're looking for good 3600-16 Kits just make sure you're finding something 3600-16-16-16 or tighter. assuming having good DDR4 Chips will matter much for you if you're only planning on adjusting the Frequency very slightly. if you can still find like a 3600 bin of Crucial BallistiX that'll be able to do very mild adjustments like that fine too and should be like half the price of a Bdie Kit.
@@taiiat0 Ya I've made my mind for ddr4 now. In 3-4 years ddr5 will easily cross 8000mhz in mid range with timings like cl30. I Think buying double expensive ddr5 5200mhz cl40 is dumb. The difference is not even 5 percent in most games. And I save money too, win win
I've been trying to understand RAM for quite a while now, but it always seems like rocket science to me, I have 2x 8GB 2400Mhz 15-15-15-35-50, trfc 421, 1.2V XMP single rank Hynix AFR RAM, you said the XMPs make your RAM performance worse, so how can I make mine better? currently I have everything set to auto except the frequency, which is set to 2933Mhz due to Ryzen 2700's maximum officially supported memory speed for 2 sticks of single rank RAM according to AMD. The auto timings are 16-21-21-49-70, trfc 514, 1.2V which I guess it's horrible, but at least it passes memtest86, is there any way to improve? I don't mind downclocking and adjusting timings if it means better performance.
GDM enable, SOC 1.15v, RAM somehwere between 1.35v to 1.4v, 16-18-18-36-56-560. tRRD/tRRDL both set to 4 and tFAW set to 16, both SCL set to 3 or 4. If stable lower CL by 2 at a time --- until crash, then lower tRCD by 1 until crash, then lower tRP by 1... then lower tRAS by 2... then set tRC to (tRP+tRAS+2) and reduce by 2 until crash. Lastly reduce tRFC by 10 until crash. should give you decent enough performance. memtest86 in today's standard is trash, it does not have (almost any) stress anymore. Use TM5 with Extreme1@anta777 profile, passing both Linpack Extreme and yCruncher is the 3 newer/better standard.
@@OMGJL Update: I've set everything like you suggested, except I couldn't find the tRRD just a tRRDS that I've set to 4 and set both SCL to 4 and set the RAM voltage to 1.35. I tried to run TM5 with the suggested profile, but it kept asking for admin privileges which I granted, but didn't make a difference, so I skipped that. I've ran linpack xtreme with the 10GB setting, which it passed then I ran the y cruncher with all test types enabled and that passed as well, so I imagine that it's stable then.
@@ryomario90 TM5 have a small bug which keeps prompt Admin thing on the beginning, not a problem. tRRDs tRRDL sometime called tRRD with no suffix, just change both the tRRD whatever to 4 and you should be good. if you can't OC to 3200 or even higher, this is about as good as it get without too much messing around.
@@OMGJL Nah... 2933 is good for me, and about the TM5 if I don't grant it admin privileges then it runs but it says in the textbox that it's running in compatibility mode and I should give it admin access, when I give it admin access then a small window will pop up when I launch the program saying AWE mode requires admin access and it pops up multiple times after I hit ' OK " then a dozen new window with errors about failing to execute the program, I don't think I'll be messing with that, but thanks for the help with my RAM OC!
CAS latencies can still be referred to as a time latency, CPU cycles are governed by time and the clock speed, so anything else based on the CPU cycles is also based on time, just with another factor
There seems to be three groups of people in this comment section, people hating on linus, people hating on you, and the small number of people who actually found this to be interesting, and are making nice friendly conversation about the topic.
I would like to ask in a simple way. Which memory is better? Kingston KF548S38IB-16 vs KVR48S40BS8-16 Both are DDR5 Both are 4800 CL40 vs CL38 Kindly reply thanks
Both are the same Clock Speed so the one who has Lower CL is Better. Kingston CL40 one: 2000"¹" × 40 ÷ 4800=16.666ns. Kingston CL38 one: 2000"¹" × 38 ÷ 4800=15.833ns. The CL38 one is better. Note"¹": we timed the CL by 2000 because we are using the Data Transfer Rate Speeds (MT/s or Mbps both is correct) that is the double of Frequency Clock (MHz) if you will use the Frequency Clock in the formula you need to Time the CL by 1000. Check the "RAM Latency Calculator Omni" on your browser for more information.
well my knowledges have increased, thank you for that, but in all fairness i never used CL as a measure of lower latency is better but rather lower CL USUALLY means a better components used on the ram; kinda like the ratings on power supplies situation. didnt know the value of it but with better understanding the ram now i know, thanks.
Sadly, where i live, in common access it is REALLY hard to get hands on any decent Samsung B-die kits... Everything there is Hynix CJR or DJR. So lower latencies for me just showing better kit, because they tend to be tightly binned.
Thank you for mentioning that memory bandwidth doesn't improve real world performance unless the application requires it (almost none do). For most applications (excluding integrated GPUs), it's very hard to differentiate between the cheapest and the most expensive ram that you can find.
It's one cycle divided by billion cycles per second (AKA GHz) not one second divided by billion cycles per second. Cycles end up in both the numerator and denominator, therefore cancel out. Seconds end up in the numerator. Billion ends up in the denominator. What you end up with is billionths of a second
question... on TR pro would dual ran mem still work better than single ranks when all 8 channels are populated? and what if you was to populate only 4 channels? *(eg for cost saving cause you don't need gargantuan amount f ram) second question.... do memory configurations eg single or dual rank or using 4 channels vs 8 available ones, or speeds speeds or latencies affect raytracing performance EG Vray RTX optix renderer. Could You get ASUS to sponsor You a TR pro memory and cpu overclocking, and see how ur RTX performance changes in VRAY bench or Octane or Blender?!
It also doesn’t help that RAM manufacturers have chosen the dumbest performance specs to advertise, because I’m sure they’re all doing it to be deliberately misleading. They insist on using mega transfers as a speed rating, because it’s DDR so it’s double the actual speed in MHz, but it really doesn’t matter if they advertise the speed as 3200M/Ts or 1600MHz, because they both mean the same thing and it’s still going to perform the same relative to something kits that are faster or slower-but I would imagine that they do it because they feel better about the bigger number since it sounds faster, right? Who cares if almost nobody knows what the heck a mega transfer is?! They shouldn’t even list the sub timings or CAS latency, because the only thing that matters is the actual latency like you’ve calculated, so it would behoove OEMs to just give that number a name and then people could shop for RAM based on that, or the fact that a kit is dual or single rank. Dual rank would likely be reflected by a lower latency score, but it’s still a worthwhile option to use for sorting the different speeds available when you’re shopping for RAM. I know that what Linus said in this instance was incorrect, but I’ve either seen you or Steve from GN explain how a significantly lower CAS latency on a kit that’s only 100Mhz/200M/Ts slower, would make it faster than the kit with higher clock speed. The issue with what Linus said is obviously because it wasn’t the case when the timings were that close, and if the price is the dramatically different you aren’t going to miss out on that slightly lower performance. Cut the guy some slack because he was live on streams and does know what he’s talking about, and he was just trying to point out that those two kits would be close enough in performance not to matter for the difference in price, so you can put that money into something that is going to give you more performance per dollar. Perhaps you could throw together a publicly shared Google Sheet with “Buildzoid’s RAM shopping calculator” where someone can just copy and paste the specs of what they’re looking at and the prices, and you can come up with some formulas to calculate the numbers people need to know in order to properly organize the data they need to make an informed purchasing decision. You can calculate a performance per dollar amount a price per GB, the amount of the actual latency of a clock cycle so they can see the raw speed numbers in a useful way. Hell, if they can list the sub timings and/or the model numbers, you could help people spot B-Die if they want to get the fastest DDR4 or if they’re getting DDR5 some SK Hynix. Seems like something you could easily get the ball rolling on and then crowd source the upkeep to your community of fans and you can just sign off on things. If you get fancy enough with the features you could probably find someone willing to develop it into a proper App you can put on a mobile device or download onto a PC. What would be neat is if you built a proper application, you could have users submit their timings and benchmarks results directly through the app to validate overclocks and custom timings and with enough user input, that could expose some patterns and give people an idea of what RAM they have and what it can do based on other user’s results, which will probably vary wildly, but might give decent baseline starting numbers to work with and increase slowly as silicon allows.
Even though I still don't have a thorough understanding of RAM, knowing a quick formula for comparing kits is very helpful. Cuts through a lot of the mystery from a buyer's standpoint.
The one tier up/one tier down thing used to be much more consistently true before DDR5, but now suddenly the frequency seems to (usually) matter much more, at least in gaming benchmarks.
Good video. Manufacturers put in thrash subtimings on XMP so that they can advertise a higher frequency/data transfer rates and make sure it's stable on a thrash motherboard with 4dimm slots/crap CPU IMC. People are stupid when it comes to RAM, they'll just look at the freq/transfer rate and CAS and buy it based on that. What I don't get though, is that there is enough room on the "SPD chip" for 2 xmp profiles and I hardly ever see 2 on a kit. Why not put in a loose and another tight timings profile? Maybe they're just afraid it's going to be unstable and get more RMA or support calls for those selecting the more aggressive timings. Doesn't matter... Xmp sucks anyway, there's too little timings in there to make a real decent tune.
Obligatory, love your content keep up the good work! But you just spent 11 minutes proving Linus right. He said that they 're basically equivalent and you agreed that anything less than 10% is hard to notice. You go on to conclude the video with the statement that the corsair ram is faster in every way, which is true, by 8% in bandwidth and 3% in latency. By your own definition they're functionally equivalent ... which is what he said! But seriously, love your work! I've used you deep dives decide which components to buy at least a dozen times now.
