AIDA's and Geekbench's memory benchmarks don't detect bad tRRD and tFAW timings
HTML-код
- Опубликовано: 6 авг 2021
- My Patreon: / buildzoid
Teespring: teespring.com/stores/actually...
The Twitch: / buildzoid
The Facebook: / actuallyhardcoreovercl...
#Overclocking #DDR4 #timings 
Наука
I know you said it’s pain, but could you please make a video about OS stripping? Of course no need to show all the secret tricks you use for world’s records competition, but at least some basics would be cool. It’s not widely spread information, so I’m really intrigued by that topic.
Also even the same benchmarks perform differently between usual installation and Benchmate package, so software quirks are very interesting issue.
luumi made a video about running CPU benchmarks on a stripped down windows XP iirc but Buildzoid's take would be cool as well
I agree, this is direly missing from this channel
@Casix03 some people would rather take the steps themselves, as the windows debloat builds often take out features that some users need/use
I actually absolutely love how buildzoid just finds these weird little random things that we all really love to learn about
Thank You Buildzoid. Your explanations here gave me a nice little chunk of education on DDR timings.
Awesome video! Now do ALL the timings in groups similarly! =X
Very good info, thank you!
Dude your a great teacher .
Thank you so much for these awesome videos
This is some top tier content. Thanks a bunch.
Thanks for spreading all yor knowledge I really appreciate your help
Damn that was a great video !
Just spent some serious change on some dual rank b-die, excited to get it into my system.
Is there a good idiots guide to setting the RTTs? Is it just brute force trialing every setting, or is there a method?
What kit did u buy?
It is primarily bashing your head against the brick wall, until either your head or the wall gives in.
Nothing exciting, just a spreadsheet bruteforcing exercise.
oh and the procODT usually determines what sort of RTT settings (e.g. Nom at /6 doesn't POST if procODT is >36.9 on my board) run in the first place, joy!
I'm so thankful for DRAM Calculator. I was able to squeeze fairly low timings out of my B-die kit without crashes in various tests & benchmarks. DRAM Calc has been a great shortcut for us Ryzen owners. I'll certainly be keeping a close eye on this channel, however, because I've learned an incredible amount of interesting things here.
Which version of Linpack to use with Zen 2 CPUs?
I tried last 2 versions of Linpack and they both report "12 cores" and "6 threads" on my 3600 during benchmark run, and therefore utilize only 6 threads.
I use linpack Xtreme from techpower up and it works fine.
Linpack is meant to be run as 1 thread/core
Hey Buildzoid, how does the Z590 Dark stack up to the Z590 Tachyon and your Z490 Apex?
What about 4 / 6 / 16 or 4 / 8 / 16? It should be more stable but is it any slower?
my machine has memory problems that I've tracked to either the mobo or CPU. Is there a software package that I can use to delve deeper into that?
Is it possible that HEDT allows for lower tFAW than 16? Because my BIOS and system info tools show that i‘m running a tFAW of 8. Also is it better to keep the two tRRD values synchronized? I’m currently running tRRD_S of 2 and a tRRD_L of 6.
tRRD minimum value is 4. it cannot be set below that. if you can set below that, it's not real.
does RRD_S and FAW relationship hold for DDR5? I have read RRD_S
What about the integrated Benchmark in the DRAM calculator from 1usmus? I find its ok to meaaure speed and stability.
I use it since im too cheap to buy aida currently lol. I finish membench in 98 seconds
Has anyone tested this with 4 6 16 timings? I keep reading that loosening tRRD_L by 2 actually improves performance. I tested using Intel MLC as well as PyPrime and I can somewhat confirm that, however the difference is so little that I can't tell whether it's just variance between tests.
Both are stable on my kit at 4000 MT/s, so what's the right setting here?
What memory benchmarks would you recommend for testing stability when tightening timings? Using Micron E die at 3600-3800. Karhu, Memtest86, HCI MemTest?
I use HCI memtest pro and Linpack
The good old free stuff unless you want to start spending money. The bootable memory test, is it HCI or is HCI the windows silly version. And google stress app test. As far as Linpack goes, it's a CPU burner, literally. There is no need to get some yet another newer Linpack benchmark as TPU download seems to suggest, just update the Linpack library in any old benchmark/stress test based on Linpack... Or run Linpack directly with what ever you want. But well, most people want clickity click and not do anything themselves. Linpack throttles pretty much all stock Intel CPUs due to power limits down to base clock, not multi core boost clock but base clock, at least when you run Linpack hard enough. Which makes it useless for testing CPU clock stability etc. even system stability of CPU clock + memory OC as the CPU downclocks and you can't test your CPU OC together with your RAM OC... Linpack is a good way to burn a CPU, literally, unless you set low power limit or the CPU is new and already has built in limiters that detect it's crazy workload and limits the CPU's performance same as GPUs limit Furmark and such to not fry themselves with these insane stress tests.
Personally found y-cruncher to be much pickier than other popular tests so I give everything a test pass through it before buckling in for anything that takes a hours. (In a similar vein, for cpu OCing I run blender bmw before the long tests)
TM5 with the anta777 extreme preset goes alright
I have 4/6/16 with micron rev e. I tried 4/4/16 it post in bios. Lowering other subtimings also works in bios. But I just finished installing a new OS and dont wont to currupt my new windows. Wish there was a way to disconnect my ssd in bios and run a seperate sata hardrive to test memory overclocking.
You can disable the controller in bios.
how u get 4.5ghz ring? u add cah voltag to what?
When i was messing with timing i just upped voltages to fix the instability. then again i only messed with primary timings ( i set 4000 cl 16-16-16-32-400 and upped the voltages from 1.4 to 1.42v ) the rest i don't understand so i left it on auto and frankly i doubt it's ideal but i just dont know how to set it and it didnt accept XMP on my 9900k rig.
Wish i could undersatand all the other timings as my copy speed just isnt good enough and it may be causing other issues i dont know of due to lack of understanding of the timings.
Not all instability can be resolved bumping voltage
If you overvolt at same frequency you gain stability only if your memory need more current or can support high overvolt
Too much overvolt without increasing frequency (adding power draw) cause instability
nice catch!
Oh, cool mobo. Hope you get to take a look at the X570 Dark in future.
I still don't really get Ring ratio/frequency. Can someone ELIF it for me?
So X570 dark when BZ?
are there even tools that adjust and test timings automatically to get the best results in 2021.
It would need to be a system written on the bios that would run tests for 24+ hours and it would still not be 100% accurate. Its a waste of time for any company to do this since any overclocker will do better manually. Just think about 30+ timings that have to run for 4hours each to check for stability, you can't simply "detect" those timings... actually, for the normies yes you can and they're called XMP, but you know what I mean.
I agree with Nuba here. Test times are just too long to make it practical. To get full lowest timings, you need at least 3+ hours of testing per timing, and there are many many timings.
Plus, you likely need dedicated hardware, as there are very few to no tools which allow you to change memory timings from within the OS. You need to edit them in BIOS, which means the only way would be to have external hardware which emulates a keyboard and can edit bios settings.
You'd also need external hardware since likely at least once for each timing, you'll end up with a system which *can't boot* and needs a CMOS reset. No way to do that from within software on the computer. Need a human, or external hardware (that there is so little profit in, no one has built)
@@daboross2 a separate system could be built for automatic tuning. There’s no need for a gpu. It can be built and it should be a thing. You just throw a ram kit in and specify your preference.
@@zazabean23 I'd like to say that's possible, but really the timings you can reach are just as dependent on your CPU's IMC and the board's trace layout as the ram itself. Yes it is largely the ram, but if the goal is 100% automation, even settings like whether the system's GPU has resizable BAR enabled or not can effect the load on the CPU's IO die enough to affect what timings you can stabilize or not.
I've personally experienced this changing motherboards, with my CPU and RAM staying identical. I went from a cheap 4-layer motherboard to a 6-layer one with a much newer memory trace design, and on my new system I could run tWR 10, where I was limited to tWR 12 on the old one. Tighter traces means better signal integrity between the CPU and RAM, which means the CPU IMC has to work less hard at decoding what's being communicated and can work faster.
The general area of what timings can be reached is definitely RAM-dependent, so I don't think your idea is completely out. It could be done. But it wouldn't be trivial to make it *actually work* with all PCs - and anything that's not running on the actual hardware is always going to require the user to then manually input those settings into their BIOS and run a full stability test themselves. And if it's then not stable, it's back to square one.
@@daboross2 100% agree, i still think there's a level of certainty that an automated RAM overclocking station could supply regarding timings and stability. I wouldn't say it's meant to give you the lowest possible timings, but more or less get you on a very good track.
The Aida Photoworxx benchmark actually responds to tFAW changes, not sure about tRRD. The score does go up significantly just from changing tFAW on both my 3900X and 5950X.
Well thaqt does explain your distain for AIDA64 software. thks
Interestingly enough, the capture card is doing perfectly fine on your Nvidia! The signal is indeed gone for a second, and then it switches to another resolution. I have a blank screen for a moment as well, and the screen has to adjust for a second. Your capture card has to switch resolutions too, that is why it says no signal. So, I mean, yeah! Good luck with the channel!
nah Nvidia cards run way worse with the capture card I'm just doing a work around to deal with the resolution changes.
@@ActuallyHardcoreOverclocking if you have not seen it already, a nice video for you BZ. ruclips.net/video/dv8TuIoSSAE/видео.html
2:09 I really, really like that wallpaper. Where can I get that?
imgur.com/gallery/cnsYnMr
If you want modify it you can get the Gimp XCF file for it from teespring
@@ActuallyHardcoreOverclocking Thank you!
Timespy/firestrike cpu tests do get affected by the gpu choice, even if it is just a small amount but it does. for instance with an 6900xt I get higher cpu test fps/scores than with my 670, but yeah it does not affect that much if you want to compare a cpu to a cpu.
Hey, I was reading about ddr4 vs ddr5, and I noticed geekbench has a huge penalty on ddr4.
you can have a 90% guess based on lower scores that its ddr4.
While other benchmarks seem to show slightly better ddr5 performance.
Aida and geekbench5 shows huge differences. Does this matter, or what is the reason?
I figured this out on my own just a month or two ago. OCCT will detect errors that hwinfo and aida won't pick up. I narrowed the culprit down to tFAW. I'm still not entirely sure why I can't lower tFAW. My motherboard is z390 Aorus Pro wifi. On auto at 3866 frequency tFAW is 48. Even if I lower it to 42 I get an error but only through OCCT's memory test. Both tRRDS and tRRDL are set to 4 and makes no difference if set higher. It's cheap Corsair RAM maybe that has something to do with it and my board being crap at memory overclocking. XMP is 3600 18-19-19-39. Maybe someone more experienced and/or more intelligent has the answer.
I tried adjusting tfaw for the first time instead of leaving it on auto and after that it started bluescreening and I ended up with a dead patriot B-die set. Thanks God we have lifetime warranty
"This could be run with a potato" so Buildzoid demos with a turnip to prove it.
So what's the best thing to use to measure deviation/gain in memory OC in the end?
Linpack
He said it many times in the video, please pay attention in the classroom.
@@N0N0111 Joke / Your Head
where can I get that wallpaper ?
imgur.com/gallery/cnsYnMr
Hey zoid man can you do a follow up on why this wrecked my ram. See my comment for details.
What a coincedence, I'm currently testing my B-Die 3200 CL16 RAM for any errors at its rated speed before trying to overclock it. Wish me luck, I'll just be reviewing your videos to see what I can take and apply to this.
You might have C-die instead.
Thaiphoon labelled it as Samsung B-Die, although I'd be open to trying other programs to get a valid identification result
@@NextEevolution Thaiphoon labels c die as b die, you will know if its c die by trying to boot 3600 cl 14 14 14 c die has bad trcd s and doesnt scale with voltage
@@hannahtimson2526 unfortunately just saw many posts of this happening. I haven't started messing around with settings yet since I'm looking at different guides. I'll have to look into how to NOT kill C-Die as well.
For starters, keep the voltage low, iirc c-die will degrade as low as 1.35v according to the ddr4 oc guide
Z590 DARK is lovely
Why not use passmark memtest86?
because I've heard from other people that it doesn't work for DDR4
@@ActuallyHardcoreOverclocking memtest86+ Doesn't work for ddr4. But Passmark very 5.0+ works for DDR4 and just added ddr5 support. Passmark was based on memtest86+ code and improved on it over time. iirc. I think many people get confused cause they both are named memtest86.
i have djr that cant do 6 6 24 at jedec 2133 clocks/timings lol
Measuring latency is how long one request takes. tRRD is about how long you have to wait between 2 requests.
Personally for me, when RAM is unstable, latency isn't constant.
After I had my settings stable, I ran the latency benchmark in Aida64 for about 10 times, and it was always the same.
DRAM calculator from 1usmus in next RAM video please 😅
I find that Aida tends to be more inconsistent right after booting into windows. If I’m benchmarking, I prefer to let a game (Could be any) with bots run along for ~10 minutes to warm everything up. Then I clear with ISLC and do my benchmarks.
I have started using MLC from intel, I find its a better tool. its a shame AIDA has barely any competition.
My AIDA issues are as follows.
1 - It takes over a minute to run memory test on my raptor lake chip but only 10 seconds on my 9900k, something odd going on there, same regardless of jedec or XMP, or tuned timings/clocks. One test took nearly 5 minutes to run, it appears to just keep looping load/idle before a result finally appears.
2 - The latency result increases by 20% on my main OS install vs bare bones test OS install (49ns vs 59ns), seems its overly affected by software running on the system. MLC on the other hand isnt. This means when people compare their results there is way too many variables in the mix on AIDA.
3 - AIDA doesnt reveal under what memory transfer size they run their latency test at, MLC does.
4 - AIDA isnt free. MLC is.
Wish everyone started testing with MLC instead. :)
"The Bank-to-Bank Delay or tRRD is a DDR timing parameter which specifies the minimum amount of time between successive ACTIVATE commands to the same DDR device, even to different internal banks." So from my point of view, this does not influence the latency of a single read/write at all, and getting the same result from a latency benchmark isn't a big surprise then.
Hmmm EVGA They are some hardcore extreme and expensive Motherboard for overclocking they are 1000 percent reliable I love their products
I think my ram broke and had to buy new ones. What do you recommend to test the new ones? I don't care about overclocking atm.
If you don’t overclock memory, then you don’t really need to test it. It just works or not.
@@funtaril You need to test new memory, no one does of course.
To test xmp, I recommend y-cruncher. Settings that worked fine in linpack, hci memtest, and well everything else, would error out with y-crunch
BTW even the timing benchmark has a margin of error OFC...
4 4 16 was stable for me, woohoo
its the double density and also now 3d nand ram modules so new software is still coming... amd 3d architecture right now
This killed my 32gb djr Dr 3600 17-21- etc kit. Won't even run jdec now.
Weird I've only had issues with 4/4/16 not being stable on some kits
So I ran HCI overnight 700%. After nuking BIOS. One error on a worker that didn't progress. Not sure if good or bad. Still jdec. Gigabyte x570 pro WiFi 1.1 with 5600x had previously run a ton of tests when I first built it. Was running 3800 fine before.
Ran 4,4,16 tfaw for awhile before I actually checked it was stable. But now I'm wondering if it's CPU.
Opened a new hci worker, to replace stuck, errors straight away 88% ram usage, jdec. All the other workers ok...
Workers 2000 each seems really sensitive to total ram usage
i noticed aida is inconsistent this year. its not worth buying. pirate it if u really want it. i dont use aida at all anymore. ANYONE PLAY WITH OCCT RECENTLY?
K thanks
Of course Aida fails... Elton John and Tim Rice should not mix their music together.
negative performance loss is performance gain, hehehe
I noticed the Aida latency test fluctuates with bclk.
Changing bclk changes memory frequency. If memory frequency changes, and timings stay the same, bandwidth and latency change.
@@smokeyninja9920 I'm talking about the way bclk rises and falls slightly on its own.
I still want him to read KERMIT THE FROG lines.
RTL IOL fix man ;)
I have a screenshot somewhere of AIDA's memory speed test showing 0MB/s from when I last tuned my RAM timings. Obviously I didn't keep those timings, but it did complete benchmarks (with terrible scores).
you just need to restart the system to fix that in my exp
I'll keep that in mind for next time :)
Who the duck uses those for memory testing lol.
People who haven't learned better
but how about games?! :S
Yes, One percent difference in 3DMark is a HUGE difference. :))))))
I was once, lifetime ago (20y+), avid OC myself too. But in the last ~7-8y I stopped step by step (fewer things) and now I only OC the GPU. I don't have the time anymore... But this itch of OC in me probably won't ever die.
Talk about DFM
It looks like AIDA really sucks
that latency is a turd for gaming
i get 36.4ns with 3600c17 (3733 14 15 30 1t all timings tuned) bdie on a maximus x apex