Tech youtubers need to STOP doing this when talking about RAM
ฝัง
- เผยแพร่เมื่อ 28 ส.ค. 2024
- My Patreon: / buildzoid
Teespring: teespring.com/...
Bandcamp: machineforscre...
Twitch: / buildzoid
Twitter: / buildzoid1
Facebook: / actuallyhardcoreovercl...
I buy DDR based on taste. The tastier ones are my favorite, obviously.
I always go for watermelon flavor :)
from my experience, kingston tastes like a fried chicken, thats why i got corsair.
RAM shaped chocolate bars when?
i prefer spicy ddr
@@TheFather_ i love chicken
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.
If they do that, nobody is gonna a DDR5 system, because the cost is so heavy for the difference in performance you are getting.
@@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.
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.
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.
There are 2 videos explaining memory timings on gamers nexus channel
@@Aggnog I watched them, but honestly didn't were more useful than some good trial and error
If you use PC Part Picker you can sort by latency!
Listening to you correct and or berate people is usually the highlight of my day.
So simply put!! Cheers.
we tend to do things like this without realizing its jealousy.
@@jrok96 Perhaps, however, I don’t feel it’s the case here. BZ has always been very adamant about accuracy.
@@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.
30 years of geeking on out computers, but this channel still makes me feel like a hapless techno winnie
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.
Idk why your not a TH-camr lol go check out Framechasers he's already doing a waay better job than these Soo called tech tubers
but 16*1/1.8 is a lot less than 40*1/2.6 (unless you're talking about general performance and not latency)
@@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.
Well it is. 3600 DDR 4 has latency of 8.88ns and 5200 ddr5 cl40 will have 15.4ns. So ddr4 is almost 2x lower cas latency. Whether it's faster will depend on the program used and how it uses memory. Although, for gaming ddr4 3600 cl16 will be way faster in general than ddr5 5200 cl40.
His target audience does not know how to activate xmp profiles.
I resemble that remark, at least until I started watching these videos
lol i was also thinking about that, but i stopped in the equivalent of "most of his audience doesn't even know what xmp is"
Most of Linus Audience is full of normies dumb dumbs.
I mean it’s true.
@@TubaWarriors you resent? Or resemble? Lol I think you meant you resent a remark like that well until you started watching those videos lol
@@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.
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.
Because when you have millions of followers and you spew nonsense many people now believe that erroneous information which is not good.
@@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. "TH-camr says something wrong omg let's all pick up our torches and pitchforks"🤦♂
That's nothing compared to the "DDR5 makes no sense, it's slower"
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.
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
Look at/study and actually timing chart... a picture is worth a thousand words. However,
You are correct : )
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.
yeah I use the speed/timing trick. But explaining to people who barely understand what timings are why that trick works is a PITA.
I used to calculate timing*2000/frequency
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.
@@eukariootti1 yes, but
15/1.8 = 8.33ns (7 button presses)
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.
@Oj10101 no you're not missing anything the difference is just under 3% which is as Linus put it "quite similar"
He literally says you go up in speed you increase latency which is not correct
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.
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
This is priceless. Thank you for pointing out a common misunderstanding. I love your educative videos!
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.
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 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.
@@bjn714 Again, I am asking about how much faster, not "which is faster"...
I think that's why a lot of companies remove the "L" and just call it CAS.
And just C** in the model number, and not CL.
more likely is just making it shorter to save space on a Sticker or something.
Or just call it CLT but I can see how that might be construed as problematic...
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?
Literally thought of you the second I saw this in the stream.
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.
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
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
@AdrianOkay no. Just no. That's just retarded to do.
@@AdrianOkay CL is actually one of the hardest timings to get stable depending on the memory sticks you're using.
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
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.
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.
Hmm
Linus tech tips is bs and drama
To be fair the guy sitting next to him is probably one of the most technologically savvy (especially on the software end) guys in all of LTT.
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.
I unsubbed because of his 3080ti video. It was fucking insulting.
Fr
@@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.
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.
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.
Anthony is silent. That's all I needed to see. 😄
"Im just gonna let him have this one.."
- Anthony, probably.
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.
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.
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.
So... The better kit has ~3% better latency? Which by your own standards is not a noticeable difference? So for a general audience, his advice was fine surely?
Exactly!
Do you have a dedicated video on how to fix/improve the secondary timings on ddr4/ddr5 with the hows and whys?
I don't think bigger TH-cam channels should "know better" when it comes to super specific topics. You don't get big on TH-cam by knowing a lot, you get big by being entertaining. You have no idea how often Linus and other youtubers get things I'm very familiar with very wrong, honestly not surprised.
The problem is that they should know better when they are giving tips. It ends giving the wrong tips, making people wasting money or not getting what they need.
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 TH-camrs use the term megatransfers more and more recently(past couple years). But the understanding usually stops there.
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.
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.
I think Linus was saying the right thing about latency. Latency can totally be measured in clock cycles, just as it can be measured in nanoseconds. Just like you can measure mass, energy, distance, velocity, etc in electron-volts, you can measure anything when you assume a standard scaling factor. You even show the unscaled latency between the two is similar, which is what I think he was referring to
What I do think he fails to mention is the overall bandwidth improvement of the 5200 kit, and the the significance of the other timings
I feel like if Linus started every technical comment he made with “In general,” he’d be better off.
He has a general knowledge of how computers work but when he reaches the limits of that knowledge on camera, you get this.
Except he was right. The difference it
Nice. Thank you for the explanation of what is going on. Now to hunt down your video for more info on the other timings .. thats somewhere..
This is why I love your channel. If I want to buy hardware I come onto your channel and see what you have to say. I even get to learn a few things!
So watching this video which goes into the technical details based on those technical details of the 15.8 vs 15.4 coming to around a 4% difference which is close that just the pure 5200 vs 4800 of 8% and listening to Linus's ground level explanation I can hereby confirm that Linus is stating things just fine, and nothing wrong with saying there's a relation to higher number CL(40 being higher than 38 numberswise) and higher speeds 5200 higher than 4800 having some form of synchronisation. 4800 with CL 36 is close to it than CL38 is sure but that wasn't something Linus was disputing, so in my opinion Linus wasn't wrong he was right, just not quite as right as you'd be if you grabbed the numbers, and that's fine.
this
Linus says 'the slower memory(referring to the 4800 kit)... although it has the lower cas latency.' and then concluding with his levels. Linus was trying to say that the lower cas latency(of the 4800 kit) is helping to mitigate the performance difference of the faster kit and thus conclude that they should perform the same. What he doesn't realise is that the faster 5200 kit is both faster and also has the lower cas latency of the two kits. Linus assumed that the lower cas latency number of the 4800 kit(38 versus 40) automatically meant that it had a lower cas latency(which is not true).
@@longdang2681 realisation is something personal, not something you can observe as another viewer. He even talked about it himself on a WAN shot with some Intel peeps. He had the realisation, just not the clarity of value's So yeah he thought CL38 on 4800 would be about CL40 at 5200, turns out that fact wasn't quite accurate enough. But the principle itself still had value, and that's why I consider what he said to be true enough, for someone not working with numbers available.
@@alleycatlordoflunes9689 Linus did not know that cas 38 on 4800 kit was a higher latency than cas 40 on 5200 kit. He was expecting the cas latency advantage of the 4800 to help mitigate the bandwidth advantage of the 5200 kit. 'Give or take' is how he structured his sentences with the 'level' thing. ie expecting the 4800 kit to lose in bandwidth but be superior in cas latency. Problem is the 4800 kit loses in both bandwidth and cas latency. This leads to his conclusion that they are 'give or take' similar in performance. He did not say that the 5200 kit is superior in both metrics but the total difference is likely not noticeable. He was wrong because he took the cas 38 at face value versus the cas 40, not because the difference between them was small(it was obvious that he was dealing with different MT/s memory). This indicates a lack of understanding of what the cas values actually mean, thus leading to the wrong conclusion.
@@alleycatlordoflunes9689 Linus : 'up a speed bin increase your latency, down a speed bin decrease your latency', never does he say up a speed bin decrease your latency. (because he's likely taking the latency at face value).A 5200 cl38 is using a lower cas latency(ns) than a 4800 cl38.
I knew this was going to happen when I heard it
You are correct except for one thing. Latency does not need to be measured in time, however I agree it would be good if it was clearer.
Linus talking out of his ass?
surprisedpikachu.jpg
Buildzoid drops the hammer 😅
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.
he is the result of what happens when you combine socially awkward nerd with the ego of a 30 year old body builder on roids
While it is nitpicky, Linus himself is nitpicky when he can say it in confidence. Fair game I’d say.
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.
And at the time of the video post, the Corsair kit had a better price too! ;-)
I get your correct in this instance, but what would your advice be to his audience? It’s probably not right to tell them just look at the speed and ignore the timings, and it’s probably wasteful to have people doing timing calcs to pick up .5ns differences in ram kits. Genuinely interesting in a simple way to explain to people how to pick a good (enough) kit, or at least be able to spot 2 kits that give the same effective performance and then you can pick based on price.
Right now DDR5 is a simple case of more speed = more better(but 6400 and higher rated kits have stability problems on many motherboards). DDR4 is simple upto 3600Mbps. Past 3600 you can get a yourself into a lot of problems if you don't know what you're doing with RAM.
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 knew something was up when i saw the double calculator pulled up
nice bz roasting LTT
Ackchyually moment
I seriously did not know that up until now. Thanks for that insight. It clears a lot of the fog around memory for me.
4:32 Minor nitpick, we are not punching in one second. It is just the reciprocal. Frequency has the dimensions "events" per second. The reciprocal has seconds per "event". If you would divide 1s you would have s²/"event".
Memory kit vendors do a really awful job about communicating their product in general
Im sure you saw his response this week. Thanks for doing this. We are all better educated because of it. Seriously, this is how we learn.
Nah, this is a long stretch. Linus as he has stated a bunch of times, tries to make tech content for people that is not nerdy, just regular people who need easy stuff to get closer to tech. So his washed out explanation is more than enough for the purpose. Everyone looking for more technical info will then go to more specific channels just as we did to end up here. But I dont think he deserves being trashed or even getting a rant about this one.
I agree entirely. The difference is 8% in bandwidth and 3% in latency. I'd qualify that as "quite similar" when addressing a crowd, which is what Linus said. If I was talking to one of my IT buddies I'd say it was marginally better. If I was talking to my dad I'd tell him they were basically the same so pick the cheaper one,
You can also use the formula CL/2000*MT/s.
Correct.
Many thanks for info on cas!
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.
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.
I have an option to buy either 3600MHz CL16 or 3200MHz CL14. When i divide the numbers, the CL14 ends up to be slightly shorter latency. So does it make sense to consider it or should i just go with the 3600MHz for some other reasons?
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.
Kingston rams usually have two profiles though.
@@kerotomas1 most don't though
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.
How can 34 people dislike this nugget of info?
Love your vids correcting the disinfo channels. Thanks BZ
5:37 LEEEROOYYYY JENKINS
So by your words 5200 CL41 IS like 4800 CL38. Also the fact that not every CPU can take high frequency being stable.
awesome, and it's in a short forma, thanks a lot
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 Thanks I'll try it, is 1.4V safe for 24/7?
@@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!
But is not this the same as Linus talked? More hz slower timing is the same as less hz faster timing?
Linus likely didn't even look at or care about the specific numbers he's been showing. The statement he formulated was a blanket one, and it could be mostly correct about latency if the numbers given happened to match up better. Differences between DDR4 3000 CL15, 3200 CL16 or 3600 CL18 fit his "neat little explanation", while DDR5 doesn't usually come with such even steps (fitting example would be something like 6000 CL40 and 6600 CL44: both at 13.3ns). Uninformed buyer won't get any use out of his offhand remark, but to be completely honest it would be useless for him to try and be specific about the memory modules he's showing. When the goal is to find a good deal on ram, your location, store and time of day will all affect the available result. It is then up to the buyer to compare the few uniquely available options and their own knowledge about timings, subtimings, ranks, voltages, etc. will be needed to make a reasonable choice. In this case I'm not sure if Linus actually did a disservice to anyone. I feel like people with some basic knowledge will see through the relation between those numbers, just like you presented in this video, and those without would still be oblivious to the minute differences and one thing being clearly the better choice. Ultimately it's good for you to make a response to such videos, because this will allow some people from the oblivious group to get informed. I deem it unlikely to affect future opinions from bigger youtubers - their formula just doesn't allow to go beyond surface level explanations on short notice, and sometimes the half-truths end up misinforming.
TLDR: Linus made a 'sometimes true' blanket statement while nonchalantly waving contrary evidence. Buildzoid hopefully managed to inform some people about what to look out for.
Agreed! The latency being
Here I am with my DDR4-3800MTs CL18-19-19 and I am sticking to that until next platform.
TL;DR: CAS Latency is measured in clock cycles, not standard units of time.
This is why I watch different youtubers and research what they're talking about.
Latency being measured in clock-cycles was the way I was taught when I was younger. Because the clock-cycle is based on a unit of time transferred into computing terms.
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).
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.
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
RAM is not the only thing these big youtubers are wrong about. If you look for PBO Guide on TH-cam, some of these youtubers give wrong information. I know this youtuber who said just do -30 all cores but but not all cores can do -30. There are more videos like these which kinda mislead people.
What's the clock rate of DDR4?
Great video...
Do you have a video explaining the mean of the other latency numbers and how it impact global perfs?
You started the data rate trend, time to start the CAS trend. No more CL timing. Only CAS. 7000C30 = 7000 CAS 30.
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.
But dont forget that you said in other video that Latency is not important; your words "THE CAS LATENCY TIMING DOESN'T MATTER AS MUCH AS YOU THINK IT DOES" so is mistake is not a big deal.
*LINUS wasn't really wrong IMO*
First of all, I'm sure he knows how it works. His frequency higher good, CL lower good approach is a good APPROXIMATION. The actual calculation AHO did worked out roughly the same (15.83 vs 15.38). The bandwidth due to frequency wouldn't be ADDITIVE to any overall latency advantage so the ACTUAL maximum synthetic performance difference is what? A couple percent?
Anyway, I really don't think the average person who listens to Linus wants or even needs the details to be nailed down in this instance. Maybe I'm wrong. (BTW, the "BLOWER" cooler analysis video from AHO was the most interesting tech video I've ever watched.)
Linus should be detailed and careful in his wording exactly BECAUSE his average viewer is a layman. Many of them base their purchases on what Linus tells them and he sort-of made it his obligation to inform his viewers well. When youre talking to someone who knows the details is when you can shrug them off, when talking to laymen however details are crucial, how else are they supposed to learn?
@@TheFlynCow You're nitpicking. This isn't advice that would cripple the performance of someone's system, especially given his audience. Bit of a dumb thing to even call him out on imo.
Thanks for teaching us!
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
I remember being annoyed about a somewhat similar thing back when LTT had the WAN show in the old house and couch and they took questions on twitter, however it was DDR3 and 1,333 MT/s vs 1,600 MT/s dual vs single channel, they just speculated without proof there wouldn't be much difference, plus single channel was a better upgrade path.
Well, it happens allot when you think the information you know is correcter, and you base your answers on it.
@@Mysteoa Then to make things worse we end up with 500 other people sreading misconceptions on message boards, Reddit, YT, YT comments etc etc which reinforces the incorrect information. Way better off having somewith the knowledge and courage to step up and say _"Whoah, that's not right. This is how it works guys/girls."_ I hate telling people things that aren't true bc what I thought has been corroborated by others I look at as _experts._
Heh, I wish he'd make this a weekly series. _"YT Channel Tech Misconceptions"_ based on videos large and small channels alike have made over the courae of the week. It would be a huge help.
Single channel Better because it was probably cheaper at the time, not better as in faster.
Take ddr5 right now it's cheap now but wait till everything uses ddr5 ram and everyone has to buy it.
The CL timings matter a lot more at lower speeds.. If you are running DDR5 at 6000Mhz, you can almost not care what the timings are.. But when you are running at DDR3 1600Mhz or DDR4 3000Mhz the times start to matter a lot more.. Bill is totally right but as we all know, the tighter the timings the fast the over all speeds will be.. Like I have FlareX 3200Mhz 14,14,14, 34, and the secondary timings are pretty tight too, But I overclocked the FlareX 3600Mhz and went with 16,16,16,36 for first set of timings and then left the secondary or Torchary times stock.. And I did manage to set the fabric to at 1800 all for a 1 to 1 and I did see close to a 200 points increase in CB-20 scores.. I averaged around 8500 with Team group 3200Mhz, Now with the overclocked FlareX to 3600, My scores went up to 8686 which was the best score I could get.. But I average 8664 on CB-20.. I have a top mounted Coolermaster 360 AIO Rad, In a Whites PHanteks P500A.. Which is a really great case for building in.. you get really good air flow which keep the temps in check.. I see people running a 5900X like me but the CB scores are much lower but I attribute all that to having different MOMBO's and the choice of the case you pick DOES MATTER.. I have a buddy with almost the same Hardware except he is in a small form factor case.. He runs about 10% slower than me.. He didn't cheap out.. He has a 3080, X570 Board, and a 5900X PBO On he tuned his timings for 3600Mhz just like me, But everything is jammed into a tinny case with barely any air flow other than the cooler's on the card and CPU.. Other than that, He has 1 120ml in the front and another 120 in the back.. He doesn't throttle but he also doesn't hit the higher single core clock speeds..
2000 x tCK / frequency (MT/s) = ns
frequency (MT/s) x ns / 2000 = tCK
Easier imo
I need “How RAM works indepth” video. Pretty please, with sugar on top.
Linus being barking up the wrong tree as usual?!? ZOMG who would have thunked?!? :P
So, the implication of this is that in general, for example, 4000 CL20 > 3600 CL18 > 3200 CL16, because they all have about the same latency but bandwidth increases with clock speed?
very roughly so, sure. while Frequency primarily affects Bandwidth it does affect everything though, which is part of the point of this Video.
yes. However on AMD Ryzen and intel 11th and 12th gen CPUs get weird above 3600 due to the memory controller having the option to run at half the speed of the memory which ruins memory latency.
@@ActuallyHardcoreOverclocking which is why myself and most others tend to target gear 1 on 12th gen (for gaming at least).
@@ActuallyHardcoreOverclocking have you ocd the 5000G Series?
My 5700G runs 4000mhz in 1:1 mode with damn nice timings - I wonder if that is a fluke or because they use a different memory controller on the apu side because of the requirements of the GPU....
?
@@justacomment1657 AMD 4000 and 5000 APUs can do 4400-4800Mbps in 1:1. They have one of the strongest DDR4 memory controllers of any CPU.