Dear Buildzoid. I have been researching which motherboard and cpu to be in my next 4 builds. (One for my dad, my best man, myself, and brother in law) and you have absolutely astounded me with your knowledge. It's surprising you don't have considerably more subscribers. When I watch other tech tubers, who I won't mention by name, they don't go anywhere near the detail you do. The first video of yours I watched was you going really in-depth on VRMs, I just watched all of your x570 videos and rather enjoyed your rants. Also, because of you, I have changed my mind on which motherboard I was going to get. I really think you should have more subscribers. I learned more watching your videos in the last few days than I have in the last 20 years of building systems. I'm just really impressed. I like how your videos don't have a lot of flash. It's substance. That matters far more to me than studio lighting and b roll of a system build. I wish you great success and look forward to more of your videos.
Good point that you can't test the IF with memory stresstests. In my case the 1900 IF passed Karhu without any problems, but playing BF V (yeah actually sometimes gaming^^) the whole system rebooted after some minutes. I had to raise voltages a little bit (Vsoc 1.15v and VDDG 1.1v) to have it completely stable. I think this video might be very helpful for some people, so thanks!
Redemz i ran Prime95 small fft few hours and lots of another benchmarks. I kind a wanna say that i got FCLK stable at 1900Mhz while UCLK is 1600Mhz. Time will tell! It gave +100 on CB20 multicore, but also gave few ns latency penalty. (x570 Taichi with 3600).
Good video. I knew CCD to CCD needed to go through IF but I didn't know CCX to another CCX on the same CCD also needed to go through IF. Makes the 3950x a bit more appealing since it has 4 cores per CCX vs 3 cores for the 3900x which could in theory be less reliant on IF if you're playing a game that uses only 4 cores, assuming Windows Scheduler is smart enough to run them on the same CCX.
This was my thoughts when the r5 1600 which is 3+3 cores and 2200g single 4core ccx first launched. (I really like the 2200g on paper and wish I had one to play with). people were praising the 1200/1300/1400/1500x which are all 2+2 cores. I'd avoid them as much as possible.
@@igoresque with hyperthreading and lots of cache because he used a 1700 and there were minor differences in favour of 4+0 but there are no quad core chips that are cheap with a potential 16MB l3 cache (even if the cache in the other CCX is disabled which we dont know thats still 8MB. Im just saying I'd just be happier about it being all in one ccx anyway.
You're right but there are no straight up quadcore games, it's mostly one pegged thread and some less strained ones. I'm just saying chances are, a game will use more than 4 threads anyway and using another CCX is faster than using SMT, but it's not so critical that *all* of the threads talk to eachother as fast as possible. I agree that schedulers have top take all this into account now though, and i hope Microsoft and other dev studios actually take up this challenge. There will be so much performance to be found if only the right cores get used at all times, instead of just round-robin.
Note that FCLK instability very often manifests itself differently than memory instability. When my FCLK was unstable I would get broken crackling audio and frequent long stutters in games without anything actually crashing. I tried pushing my SoC voltage up to 1.2 V and couldn't make 1900 MHz stable. I think it was actually getting worse at some point. I still seem to manage 1866 MHz at 1.12 V or so quite easily.
@Actually Hardcore Overclocking The Video was quite good but your diagram is not quite accurate. There is only a single interconnect that connects each of the CCds to the IO die and that Interconnect is chared by both CCX modules. There are two interconnects for the two die CPUs 3900x and 3950x. Each of the Memory related interconnects operate at a rate of FCLK x 32 Bytes/cycle. You can work out the max bandwidth by calculating it out. with 3600MT/s memory the total available bandwidth is 57.6GB/s. That is teh reason why the max throughput that you observed was 57GB/s even with higher clocked memory. Intel Ringbus runs the connectivity between CPU and memory at a badwidth determined by teh Cache multiplier. At x45 that is 144GB/s. Both the Infinity Fabric and Ringbus provide shared bandwidth for teh CPU accessing memory and the GPU accessing the L3 Cache to obtain instructions. Tested memory bandwidth on Intel remains constant and in excess of the maxumin throughput potential of teh DDR4 memory but with Ryzen, the tested throughput is only true when you are not running a game as the Interconnects Bandwidth at 1:1 matches the capacity of the Installed Ram. Running a GPU in a game over the shared bandwidth means that the CPU bandwidth to memory is compromised. As such, the CPU obtain data necessary to operate at full capacity. As you alluded to, a system with 2133 MT/s Ram with 1900FCLK will almost match a system with 2080TI running 3800MT/s Ram when a 3200 system will not in games.
Thanks buildzoid I had been discovering all this for my self the stability issues on full load and I came to the same conclusion cheers bud u smartie pants =)
Didn't think I would understand any of this. but thanks to the avoidance of all the jargon combined with easy to understand abbreviations, I actually learnt quite a bit, Great job on the video.
G'day Buildzoid, I like the videos that focus on 1 topic, I've not been understanding the Core connections + Memory +Infinity Fabric & abbreviations but you picture & explanation just made it really simple, Thanks, I'm saving this in my favourites
This cleared up a few things I wasn't sure about, so thanks a lot for this type of content. I know it's relatively niche stuff, but still useful to know. Even for those of us who aren't ACTUALLY hardcore overclockers, but still have an interest in the nitty gritty stuff.
Your video's have learned met a lot over the last few weeks, I'm happy to have found your channel (y). You say in the beginning of this video that the memory controller is a mess but I haven't seen any proof of that in the video because it's just .... a little complicated haha :) Running a 3900X over here with G.Skill NEO (16-19-19-19-39-3600) 2 x 16 DRAM over here which works fine but doesn't seem to like overclocks so I'm quite happily running stock on CPU/DRAM/GPU(1080ti). Thank you for your great videos /me like !
I understand that MCLK is DDR4 over 2 and I understand the UCLK clocks the UMC and the UMC branches off to the CCD and then to a CCX, but none of this tells us why SM can't be in the MCU anymore? \o/ ;P
Is that Samsung B-Die or Hynix CJR? I bought G.Skill Ripjaws V 3200MHz (f4-3200c16d-16gvkb) for $77.99 (Hynix CJR) and I'm really pleased with them. They ran 3400 at 14-16-16-16 1.40v and now run 3600MHz 16-19-19-19 1.38v. I could probably tighten a few subtimings (especially TRFC) but I'm already happy with the increase in responsiveness from 1600 to 1800 FCLK and I want to enjoy it and stop tweaking stuff :D
Thanks for the amazing video. I now have graduated in how to run an amd CPU... Honestly you can tell how everyone was lost at the release. I wanted to buy a 3900x but when I saw what was going on on userbench, I'm just gonna wait for intel 10th gen and go from here. I'm not an intel fan, I'm glad to see amd come back to improve competition but man at least when you run a new intel cpu you know what you get right out of the box without marketing spiel.
very nice explained -as a follow up it would be interesting to see those settings directly in the bios... ( one time with cheap ram 3000 - 3200 mt and one time with 3800 mt ) :D
Could you please do a video about the performance of the 32gig samsung oem stick you showed us in a previous video and test how these oc on x570 platform in a 2 (64G) and 4 (128G) sticks setup?
It would be neat if the 32 core TR could still USE 8 channels on a new motherboard while maintaining backwards compatibility with the previous 4 channel boards. Not sure how big the performance difference going from 4 to 8 channels would be for a 32 core, but might not be worth it if those boards are like 1k.
Casmar 8 channels is useful for workloads like database and big data but the price hike definitely will impact the regular consumer. I think AMD will reserve 8 channels for Rome.
I still often wonder how the latency and overall performance would be with 4000 MT RAM, 2000 mhz uclk 2000 mhz fclk and 4000 mhz fixed core clock. Would probably take LN2 to reach those uclk and fclk if it's even possible, but it'd be neat to see what a 2:1 core to everything else would do.
Since the IO die can be binned as well, I wonder if we'll see over 1900MHz FCLK stable on the 3950x. From what I've seen only the 3600 has a problem getting 1800MHz, so they probably get all the dud IO dies.
My dyslexia hates this video hahaha Thanks for all your very informative videos! They have helped me with my first pc build in a very long time, of which I will have all the parts tomorrow lol
I've never learned so much yet understand almost nothing before. Really good video, I'll definitely have to rewatch this multiple times to actually understand though lol
I recently discovered the hard way that more SoC voltage is not the way to get on zen2. I was playing around with a friends 3600, pushed IF to 1900 at 1.11v (1.1v GET), and was getting WHEA errors. Tried pushing more SoC and things just got worse and worse, at 1.16v (1.15v GET) things like video playback would chug and audio would crackle, and there were far more WHEA errors. I was so confused, dropped SoC to 1.09v and all the issues disappeared and things were stable. I had never seen this issue before and I've messed around with 9 or 10 different ryzen systems, only 4 or 5 were zen2, and never saw issues up to 1.2v SoC on zen/+.
HUB did a video on the ratios of UCLK and MCLK and 1:1 is definitely better. I recall performance gains from 10-20% from going from 3200 stock timings ram to up around 3700 with all the timing pulled tight. I recall OCing the UCLK more than the exact 1:1 resulted in a loss as well.
Question: people seem to think that you always need to run 1:1 on Ryzen 3000, but from what i've seen it is beneficial to overclock IF (FCLK i believe) even if it will no longer run 1:1 with RAM. The only way non-1:1 becomes a problem is if you are running RAM past 3800Mhz in 2:1 mode. So for example pushing FCLK to 1900 but MCLK is at 1600 (3200Mhz RAM). By this logic people seem to think that you should also run FCLK at 1600 so it would be 1:1 Hope someone can clear this up for me.
o snap, i think u solved my issue, i was running my 3200 mhz kit at 3600 mhz (sorry mega ticks? kappa ) and 1900 fclk, so even though memtest came out 100% error free when i would encode long videos or game for a long time i would crash, so i was over stressing the infinity fabric and was also against your 166 rule, so i should keep it fclk = uclk so both at 1800, i am gonna try that right now, thanks bro
i'm Commander Shepard and this is my favorite Ryzen video on the Citadel. (well maybe, idunno but either way yes, this video is useful, and yes, it is interesting. no need to worry about that)
Great work bud. 👍 Gonna be fun fine tuning mine when I get it. 😁 Wendel from Level1Tech has been on the case too and he's found some interesting info as well. 😉
Maybe im dumb but what exactly was the reason to be above 166 difference between FCLK and UCLK? AMD says one sweetspot is 3600 mhz ddr4 1:1, that wont get you 166 mhz difference between FCLK and UCLK.
So, on the Ryzen 7/9 DDR 3600 is ideal. MCLK, UCLK, and FCLK is 1800MHz. I have DDR 3200 with Samsung DIE's, so hopefully I can get it from 1600 to 1800.
Question: my initial thought was that the FCLK worked best when equal to MCLK ie. 3600 kit to give 1800Mt/s and clock the infinity fabric (fclk) to 1800. You suggest this is not optimal and performance can be improved by having an offset of 166 between the two. So if you have an fclk of 1800 you need a uclk
How fast? More like how slow and hot. Everybody used to do it that way, but those crossbars ran hot, limiting performance (remember, most of the market at large cares about actual TDP), so mesh type interconnects and ring busses are the order of the day.
About to pull the trigger on a 3600 and pair it with some gskill ripjaws 2x8gb sticks, 3400 cl 16 on the updated msi b450 tomahawk max. What memory profile do you recommend? I guess it will depend on how far my cpu and ram can go (and btw whats you opinion of this ram in particular, I read in foruns it can be sort of a gamble as you may get a good samsung die or really shitty micron one). Thank you bz, love your content 👍
The real clockspeed of DDR4 3200"MHz" is 200 real MHz which is doubled 4 times... therefore the latenzy did not changed since I dont know 10 years... it is simply a bandwith increase...
Excellent video as aways. Iam surprised how dificult is to get correct information this days. That information about inter-ccx communication is really hard to find. please, tell us, where did you get it?
Hey Buildzoid, nice job! Can you make another similar video explaining ryzen 3000 memory system with diffrent memory configuration - single/dual rank, two sticks, four sticks, 2x8gb, 4x8gb etc? Some pros and cons, teoretical ideal setup.... Thank you :)
I've read 2:1 mode doesn't make the UCLK run at 50% of MCLK, but actually clamp it at 1800Mhz while letting MCLK go higher. Otherwise going in 2:1 would have tremendous latency, far beyond a simple out-of-sync package. I've bought a 4x16Gb 3200Mhz kit that simply doesn't want to go over 3000Mhz :(
14:05 - lol I can relate because I have some Micron D9TGG that doesnt OC for shit and some Hynix BJR that only gets to 3066Mhz and no more where as my hynix AFR tops out at 3800mhz, fun fact: these 4 dimms with 3 different ICs (2 being AFR) is the same 4X4GB Kit of kingston HyperX 2400Mhz (DDR4)
I have one of those rare 3000-only memory kits unfortunately, a set of four Corsair Dominator Platinum C15 8GB chips with SK Hynix dies. I tried reeeeally hard to get 3200 out of them, even in the range of 3100, and just absolutely could not achieve stability lol
Hi I would like to make a question: I have bought these RAM sticks "G.Skill TridentZ Neo 16GB DDR4-3600MHz (F4-3600C16D-16GTZNC)" because they were way cheap and I didn't know they were Hynix RAM sticks. I do NOT intend to OC them, I only OC the CPU. What you suggest I do? Keep them or buy "G.Skill TridentZ Neo 16GB DDR4-3600MHz (F4-3600C16D-16GTZN)" RAM? I really trust your opinion and would really love your advise. I have an Asus Rog Crosshair VII Hero MoBo combined with a 3600 Ryzen 5 CPU. Best
can you do a video about epyc too? my company plans to buy new servers at the end of the year if dell delivers... so I would be really interested in that. :)
Welcome home, and i was wondering how slow ram needs to be before more fclk = better thanks. If you get a chance can you test your x470 gaming plus on a ryzen 3000 cpu, i can't pass memtest86+ 5.01 (fails in test 7) without either using single channel ram or disabling cores, i tested 2 ram kits w/ same results, not sure if it is the bios or a bad cpu, and if i turn off SMT and have all cores it locks up 1/2 way through test 7 in under 4 minutes; i have a 3600
I bought £125 3600 C19-20-20-40 2x16gb (alleged to be hynix cjr) for my 3900x/B450 Tomahawk Max. Set XMP, set DRAM Voltage to 1.35 manually. Works fine. Aparently there is some headroom for OC/latency tweaks but i've yet to get around to it
@@punktkomma9489 it's actually not that bad if you calculate the delay in clocks.... and the ram was so cheap, like 33% price for maybe 3% perf loss... but yes i always want better, it'll save me from being tempted to get some of that gskill 3800 CL14 lol
Would running MCLK/FCLK at some other, simple-ish fraction help in non 1/1 cases over picking some random number, say 5/4 or 6/5 (in your example 1760 or 1833MHz)?
@Actually Hardcore Overclocking - was it just a simplification, or does the memory layout really mix a daisy chain and t-shape layout for each channel? If so, wouldn't it have some negative impact on stability and performance?
does this mean that using somehing like 2400 DDR4 memory kit should be viable as well? MCLK =1200 Mhz , so if UCLK would be able to run at 1200 Mhz, therefore at 1:1 ratio. According to your rule of FCLK-UCLK=>166 , you would be getting with the result of this equation somewhere in the 700, even with crappy CPU which can do FCLK 1766. Which in theory means, you should be able to negate the silicon lottery, because you just dont care if the chip is good or bad. Can someone please explain if this is possible and what effects on performance this would have?
I need to go out more....I understood every word
Get your favorite gender to fuck you into a vegetable, should solve that :P
did you understand the mistake he made too?
XD
I appreciate those who count from 0 instead of 1.
Happy zero noices
counting from "0" is so yesterday... I now count from -1, use the inverse and then subtract 1 as my starting point.
@@photonboy999~0xFFFF + 0xFFFF? Why would you do this or count like this? Just use 0x0.
Because there are only 10 types of people in the world?
Yep, the way I learned in the late 70s till now.
Thank you so much - those kind of Videos really rock .
You should do them more often, you are really doing great explaining that kind of stuff!
Dear Buildzoid.
I have been researching which motherboard and cpu to be in my next 4 builds. (One for my dad, my best man, myself, and brother in law) and you have absolutely astounded me with your knowledge.
It's surprising you don't have considerably more subscribers.
When I watch other tech tubers, who I won't mention by name, they don't go anywhere near the detail you do. The first video of yours I watched was you going really in-depth on VRMs, I just watched all of your x570 videos and rather enjoyed your rants.
Also, because of you, I have changed my mind on which motherboard I was going to get.
I really think you should have more subscribers. I learned more watching your videos in the last few days than I have in the last 20 years of building systems.
I'm just really impressed. I like how your videos don't have a lot of flash. It's substance. That matters far more to me than studio lighting and b roll of a system build.
I wish you great success and look forward to more of your videos.
Good point that you can't test the IF with memory stresstests.
In my case the 1900 IF passed Karhu without any problems, but playing BF V (yeah actually sometimes gaming^^) the whole system rebooted after some minutes. I had to raise voltages a little bit (Vsoc 1.15v and VDDG 1.1v) to have it completely stable.
I think this video might be very helpful for some people, so thanks!
Flynnchen82 prime95 does a good job testing IF
@@rdmz135 Which prime95 mode?
@@cl4ster17 I do both blended and small FFT. Both do the job but blend is probably enough to test the IF.
@@rdmz135 Good to know. Thanks
Redemz i ran Prime95 small fft few hours and lots of another benchmarks. I kind a wanna say that i got FCLK stable at 1900Mhz while UCLK is 1600Mhz. Time will tell! It gave +100 on CB20 multicore, but also gave few ns latency penalty. (x570 Taichi with 3600).
Good video. I knew CCD to CCD needed to go through IF but I didn't know CCX to another CCX on the same CCD also needed to go through IF. Makes the 3950x a bit more appealing since it has 4 cores per CCX vs 3 cores for the 3900x which could in theory be less reliant on IF if you're playing a game that uses only 4 cores, assuming Windows Scheduler is smart enough to run them on the same CCX.
This was my thoughts when the r5 1600 which is 3+3 cores and 2200g single 4core ccx first launched. (I really like the 2200g on paper and wish I had one to play with).
people were praising the 1200/1300/1400/1500x which are all 2+2 cores. I'd avoid them as much as possible.
@@tomstech4390 Steve from Hardware Unboxed had tested this and saw no actual difference between 2+2 and 4+0 in gaming.
@@igoresque with hyperthreading and lots of cache because he used a 1700 and there were minor differences in favour of 4+0 but there are no quad core chips that are cheap with a potential 16MB l3 cache (even if the cache in the other CCX is disabled which we dont know thats still 8MB.
Im just saying I'd just be happier about it being all in one ccx anyway.
You're right but there are no straight up quadcore games, it's mostly one pegged thread and some less strained ones. I'm just saying chances are, a game will use more than 4 threads anyway and using another CCX is faster than using SMT, but it's not so critical that *all* of the threads talk to eachother as fast as possible.
I agree that schedulers have top take all this into account now though, and i hope Microsoft and other dev studios actually take up this challenge. There will be so much performance to be found if only the right cores get used at all times, instead of just round-robin.
if AMD will use interposers for Zen 3 they probably can get much higher FCLK (because silicon will clock higher than substrate connections)
I'm from the future, they used the same umc
Literally the only video I know that explained Ryzens memory system perfectly understandable . Highly appreciated and liked.
Note that FCLK instability very often manifests itself differently than memory instability. When my FCLK was unstable I would get broken crackling audio and frequent long stutters in games without anything actually crashing. I tried pushing my SoC voltage up to 1.2 V and couldn't make 1900 MHz stable. I think it was actually getting worse at some point. I still seem to manage 1866 MHz at 1.12 V or so quite easily.
I was wondering what that was, settled on 3200 cl16 with 1600 IF
@Actually Hardcore Overclocking The Video was quite good but your diagram is not quite accurate. There is only a single interconnect that connects each of the CCds to the IO die and that Interconnect is chared by both CCX modules. There are two interconnects for the two die CPUs 3900x and 3950x.
Each of the Memory related interconnects operate at a rate of FCLK x 32 Bytes/cycle. You can work out the max bandwidth by calculating it out. with 3600MT/s memory the total available bandwidth is 57.6GB/s. That is teh reason why the max throughput that you observed was 57GB/s even with higher clocked memory.
Intel Ringbus runs the connectivity between CPU and memory at a badwidth determined by teh Cache multiplier. At x45 that is 144GB/s. Both the Infinity Fabric and Ringbus provide shared bandwidth for teh CPU accessing memory and the GPU accessing the L3 Cache to obtain instructions.
Tested memory bandwidth on Intel remains constant and in excess of the maxumin throughput potential of teh DDR4 memory but with Ryzen, the tested throughput is only true when you are not running a game as the Interconnects Bandwidth at 1:1 matches the capacity of the Installed Ram. Running a GPU in a game over the shared bandwidth means that the CPU bandwidth to memory is compromised. As such, the CPU obtain data necessary to operate at full capacity. As you alluded to, a system with 2133 MT/s Ram with 1900FCLK will almost match a system with 2080TI running 3800MT/s Ram when a 3200 system will not in games.
wow i never knew kermit knew so much about tech
I bought the 4400 cl19 "(really good b-ie") Ram that you recommended in another video, and I have had some very good results from the sticks.
Thanks buildzoid I had been discovering all this for my self the stability issues on full load and I came to the same conclusion cheers bud u smartie pants =)
I bought an AHOC VRM shirt via Teespring and it is high quality and it even comes in rare tall sizes. Thanks for your content and information BZ!
You are the man! Thank you for share those valuable informations with us ;) I really wondered them.
Didn't think I would understand any of this. but thanks to the avoidance of all the jargon combined with easy to understand abbreviations, I actually learnt quite a bit, Great job on the video.
the memory math reminds me of the good ole athlon overclocking days, ah nostalgia.
Another awesome video as usual buildzoid! Im glad you're back!
G'day Buildzoid,
I like the videos that focus on 1 topic, I've not been understanding the Core connections + Memory +Infinity Fabric & abbreviations but you picture & explanation just made it really simple, Thanks,
I'm saving this in my favourites
This cleared up a few things I wasn't sure about, so thanks a lot for this type of content. I know it's relatively niche stuff, but still useful to know. Even for those of us who aren't ACTUALLY hardcore overclockers, but still have an interest in the nitty gritty stuff.
Your video's have learned met a lot over the last few weeks, I'm happy to have found your channel (y). You say in the beginning of this video that the memory controller is a mess but I haven't seen any proof of that in the video because it's just .... a little complicated haha :) Running a 3900X over here with G.Skill NEO (16-19-19-19-39-3600) 2 x 16 DRAM over here which works fine but doesn't seem to like overclocks so I'm quite happily running stock on CPU/DRAM/GPU(1080ti). Thank you for your great videos /me like !
BTW ... my prediction ... ZEN 3 will have no more cores but will have 8-core per CCD .... We'll get back to this next year :P ;)
@@b1lleman Zen 2 already has 8 cores per CCD.
Thank you! This helped me optimize my Ryzen 3900X! :)
This video is very good, i plan on overclocking my ram and this has helped me understand how all of this works
I understand that MCLK is DDR4 over 2 and I understand the UCLK clocks the UMC and the UMC branches off to the CCD and then to a CCX, but none of this tells us why SM can't be in the MCU anymore? \o/ ;P
because Disney & Sony are Selfish 😒 I'd rater Spidey than Captain Marvel
What the heck is SM?
@@AstralS7orm Spider-man in Marvel Cinematic Universe 😁
@@AstralS7orm Spider-Man ;)
Got my Trident Z RGB 3200MHz CL16 OC'd to 3600MHz CL16. So yeahh....
Is that Samsung B-Die or Hynix CJR? I bought G.Skill Ripjaws V 3200MHz (f4-3200c16d-16gvkb) for $77.99 (Hynix CJR) and I'm really pleased with them. They ran 3400 at 14-16-16-16 1.40v and now run 3600MHz 16-19-19-19 1.38v. I could probably tighten a few subtimings (especially TRFC) but I'm already happy with the increase in responsiveness from 1600 to 1800 FCLK and I want to enjoy it and stop tweaking stuff :D
Did you do a video on that topic regarding zen and zen+ ?
Awesome content as always AHC.
There wasn't much to explain, previous Zens and even older families were a single die, shared most of the cache, etc.
@@jacekjagosz Thanks, I see. So, well there is. But the stuff can be looked up.
Since DDR stands for Double Data Rate, it is correct to call it 4000MHz when the MCLK is really 2000MHz. Because that is double the data rate.
/s
Thanks for the amazing video. I now have graduated in how to run an amd CPU...
Honestly you can tell how everyone was lost at the release. I wanted to buy a 3900x but when I saw what was going on on userbench, I'm just gonna wait for intel 10th gen and go from here.
I'm not an intel fan, I'm glad to see amd come back to improve competition but man at least when you run a new intel cpu you know what you get right out of the box without marketing spiel.
Great, i understand now why GCC was unstable but memtest was stable. Thx for this
very nice explained -as a follow up it would be interesting to see those settings directly in the bios... ( one time with cheap ram 3000 - 3200 mt and one time with 3800 mt ) :D
Wow! I was blind now i can see! Thank you for this perfect explanation.
Man, this guy is using GIMP, he could have used Adobe "Intel" optimized Photoshop but he is using GIMP
Photoshop runs just fine on Ryzen. It opens in like 2 seconds, filters run fast... I guess cores > optimization.
Could you please do a video about the performance of the 32gig samsung oem stick you showed us in a previous video and test how these oc on x570 platform in a 2 (64G) and 4 (128G) sticks setup?
This video was unusually good quality.
Really interesting topic aswell!
Keep up the good vids. :)
There might be an octa-channel TR variant in the works, but that will require new motherboards.
It would be neat if the 32 core TR could still USE 8 channels on a new motherboard while maintaining backwards compatibility with the previous 4 channel boards. Not sure how big the performance difference going from 4 to 8 channels would be for a 32 core, but might not be worth it if those boards are like 1k.
@@NeoNoggie the socket will stay the same, so there is a possibility
Casmar 8 channels is useful for workloads like database and big data but the price hike definitely will impact the regular consumer. I think AMD will reserve 8 channels for Rome.
Thank you for this video! Really well done.
makes so much sense for how to optimizing parallel programs for ryzen, thanks
Really great video, explaining a lot!
Thanks!
yea, i stopped running 1900 fclk myself because it was too borderline stable.
I still often wonder how the latency and overall performance would be with 4000 MT RAM, 2000 mhz uclk 2000 mhz fclk and 4000 mhz fixed core clock. Would probably take LN2 to reach those uclk and fclk if it's even possible, but it'd be neat to see what a 2:1 core to everything else would do.
Since the IO die can be binned as well, I wonder if we'll see over 1900MHz FCLK stable on the 3950x. From what I've seen only the 3600 has a problem getting 1800MHz, so they probably get all the dud IO dies.
Thanks for the information. Good to know and it makes sense to me.
So the CLUCK has to run twice as fast as the FUCKLER...got it! Thanks.
😏
From this it looks reasonable for Zen4 to have 8 cores per CCX and 2 CCX per CCD.
Got to keep the latencies in check when scaling up the core count.
Amazing video, thanks for the information.
Hope to see you redo this video for Ryzen 7000 and with DDR5 and such, even if it's essentially the same info
Brilliant explanation in laymans terms Buildzoid 👌
My dyslexia hates this video hahaha Thanks for all your very informative videos! They have helped me with my first pc build in a very long time, of which I will have all the parts tomorrow lol
I've never learned so much yet understand almost nothing before. Really good video, I'll definitely have to rewatch this multiple times to actually understand though lol
So why use 3600mhz memory when most ryzen processors support 3200mhz?
I recently discovered the hard way that more SoC voltage is not the way to get on zen2. I was playing around with a friends 3600, pushed IF to 1900 at 1.11v (1.1v GET), and was getting WHEA errors. Tried pushing more SoC and things just got worse and worse, at 1.16v (1.15v GET) things like video playback would chug and audio would crackle, and there were far more WHEA errors. I was so confused, dropped SoC to 1.09v and all the issues disappeared and things were stable. I had never seen this issue before and I've messed around with 9 or 10 different ryzen systems, only 4 or 5 were zen2, and never saw issues up to 1.2v SoC on zen/+.
HUB did a video on the ratios of UCLK and MCLK and 1:1 is definitely better. I recall performance gains from 10-20% from going from 3200 stock timings ram to up around 3700 with all the timing pulled tight. I recall OCing the UCLK more than the exact 1:1 resulted in a loss as well.
Question: people seem to think that you always need to run 1:1 on Ryzen 3000, but from what i've seen it is beneficial to overclock IF (FCLK i believe) even if it will no longer run 1:1 with RAM.
The only way non-1:1 becomes a problem is if you are running RAM past 3800Mhz in 2:1 mode.
So for example pushing FCLK to 1900 but MCLK is at 1600 (3200Mhz RAM). By this logic people seem to think that you should also run FCLK at 1600 so it would be 1:1
Hope someone can clear this up for me.
Ok so this is addressed at 12:28 onwards in the video i believe.
U can check biuldzoid's video called 'Ryzen 3rd gen MCLK vs FCLK part 2'. At the beginning of this video he talks about this case scenario.
Yeah watching the video first is usually a good idea.
The prodigal son returns! About damn time 😁 This is going to help me when I get my Ryzen 5 3600 and 16GB (2x8) of 3200 MHz Micron E-Die
o snap, i think u solved my issue, i was running my 3200 mhz kit at 3600 mhz (sorry mega ticks? kappa ) and 1900 fclk, so even though memtest came out 100% error free when i would encode long videos or game for a long time i would crash, so i was over stressing the infinity fabric and was also against your 166 rule, so i should keep it fclk = uclk so both at 1800, i am gonna try that right now, thanks bro
Feels good to be back.
i'm Commander Shepard and this is my favorite Ryzen video on the Citadel.
(well maybe, idunno but either way yes, this video is useful, and yes, it is interesting. no need to worry about that)
Great work bud. 👍
Gonna be fun fine tuning mine when I get it. 😁
Wendel from Level1Tech has been on the case too and he's found some interesting info as well. 😉
Great video as usual Thank You!
Well they need to get there caches and branches to be better. That will remove some of that issues
Excellent explanation that is really helpful thanks
Maybe im dumb but what exactly was the reason to be above 166 difference between FCLK and UCLK? AMD says one sweetspot is 3600 mhz ddr4 1:1, that wont get you 166 mhz difference between FCLK and UCLK.
So, on the Ryzen 7/9 DDR 3600 is ideal. MCLK, UCLK, and FCLK is 1800MHz. I have DDR 3200 with Samsung DIE's, so hopefully I can get it from 1600 to 1800.
Welcome back :)
Question: my initial thought was that the FCLK worked best when equal to MCLK ie. 3600 kit to give 1800Mt/s and clock the infinity fabric (fclk) to 1800. You suggest this is not optimal and performance can be improved by having an offset of 166 between the two. So if you have an fclk of 1800 you need a uclk
This is very helpful. I can't wait to break my future ryzen 3k!
If I didn't understand this video inherently wrong, I should be good with 3600MHz CL14 and FCLK 1800MHz
makes you wonder how fast it would be if all cores could access each other without the need of infinity fabric.
I wouldn't expect anything crazy until at least AM5. That will probably let them ditch the external chipset too, more pins is more better.
How fast? More like how slow and hot. Everybody used to do it that way, but those crossbars ran hot, limiting performance (remember, most of the market at large cares about actual TDP), so mesh type interconnects and ring busses are the order of the day.
does FCLK scale with LN2?
There's no reason why it shouldn't, though it would scale differently.
About to pull the trigger on a 3600 and pair it with some gskill ripjaws 2x8gb sticks, 3400 cl 16 on the updated msi b450 tomahawk max. What memory profile do you recommend? I guess it will depend on how far my cpu and ram can go (and btw whats you opinion of this ram in particular, I read in foruns it can be sort of a gamble as you may get a good samsung die or really shitty micron one).
Thank you bz, love your content 👍
In my case 1900 FCLK stable (Ryzen5 3600) pass test AIDA64, linx, "AAA Game" about 1 hour all goes well. (RAM 3200cl14 SOC voltage 0.975)
Games rarely stress IF enough either. Try a video encode as mentioned.
The real clockspeed of DDR4 3200"MHz" is 200 real MHz which is doubled 4 times... therefore the latenzy did not changed since I dont know 10 years... it is simply a bandwith increase...
Excellent video as aways. Iam surprised how dificult is to get correct information this days. That information about inter-ccx communication is really hard to find. please, tell us, where did you get it?
why not call this Ryzen 3000 memory system ramblings? :D
Hey Buildzoid, nice job! Can you make another similar video explaining ryzen 3000 memory system with diffrent memory configuration - single/dual rank, two sticks, four sticks, 2x8gb, 4x8gb etc? Some pros and cons, teoretical ideal setup.... Thank you :)
My FCLK topped out at 1866 feelsbad. 99% sure my memory would OC further too :( Can't even post to bios with 1900 FCLK I was so sad.
I've read 2:1 mode doesn't make the UCLK run at 50% of MCLK, but actually clamp it at 1800Mhz while letting MCLK go higher.
Otherwise going in 2:1 would have tremendous latency, far beyond a simple out-of-sync package.
I've bought a 4x16Gb 3200Mhz kit that simply doesn't want to go over 3000Mhz :(
That's not true at all. 1:1 and 2:1 are the only options. It can't run decoupled.
@Carpelus Not so far. Kinda annoying because those are TridenZ Royal. I was expecting them to be at least be able to run at stock speed.
Thanks a lot, that was helpful!
14:05 - lol I can relate because I have some Micron D9TGG that doesnt OC for shit and some Hynix BJR that only gets to 3066Mhz and no more where as my hynix AFR tops out at 3800mhz, fun fact: these 4 dimms with 3 different ICs (2 being AFR) is the same 4X4GB Kit of kingston HyperX 2400Mhz (DDR4)
I have one of those rare 3000-only memory kits unfortunately, a set of four Corsair Dominator Platinum C15 8GB chips with SK Hynix dies. I tried reeeeally hard to get 3200 out of them, even in the range of 3100, and just absolutely could not achieve stability lol
I only clicked on this because I saw a creeper in the thumbnail.
What I'm interesting to know is if my PC will work with mixing 2x8GB DDR 4 at 3600MHz CL18 and 2x8GB DDR 4 3200MHz CL16?
On 16:00 - 1833 is 3666MT/s for DDR4 and not 3533.
Oh, and Welcome Back! Gimp and I missed you.
Hey i have a 980ti and 970 that may still work but as of now do not any chance you would want them for free?
Could have drawn a Crossbar (Xbar) in the IO - die :)
Hi I would like to make a question: I have bought these RAM sticks "G.Skill TridentZ Neo 16GB DDR4-3600MHz (F4-3600C16D-16GTZNC)" because they were way cheap and I didn't know they were Hynix RAM sticks. I do NOT intend to OC them, I only OC the CPU. What you suggest I do? Keep them or buy "G.Skill TridentZ Neo 16GB DDR4-3600MHz (F4-3600C16D-16GTZN)" RAM? I really trust your opinion and would really love your advise. I have an Asus Rog Crosshair VII Hero MoBo combined with a 3600 Ryzen 5 CPU. Best
can you do a video about epyc too? my company plans to buy new servers at the end of the year if dell delivers... so I would be really interested in that. :)
Wow this video was super cool. Where did you learn all this? Also, what would you recommend for stress testing the entire system? P95 small FFTs?
Small FFT doesn't really test memory. You'd want blend or something like AIDA64 stress test.
Very useful information.
Welcome home, and i was wondering how slow ram needs to be before more fclk = better thanks. If you get a chance can you test your x470 gaming plus on a ryzen 3000 cpu, i can't pass memtest86+ 5.01 (fails in test 7) without either using single channel ram or disabling cores, i tested 2 ram kits w/ same results, not sure if it is the bios or a bad cpu, and if i turn off SMT and have all cores it locks up 1/2 way through test 7 in under 4 minutes; i have a 3600
Welcome back
75k subs.... damn thats alot!
I bought £125 3600 C19-20-20-40 2x16gb (alleged to be hynix cjr) for my 3900x/B450 Tomahawk Max. Set XMP, set DRAM Voltage to 1.35 manually. Works fine. Aparently there is some headroom for OC/latency tweaks but i've yet to get around to it
Yeah you definitely wanna get the timings down, CL19 on 3600MHz is horrible. May need higher voltages to get a lower cas latency
@@punktkomma9489 it's actually not that bad if you calculate the delay in clocks.... and the ram was so cheap, like 33% price for maybe 3% perf loss... but yes i always want better, it'll save me from being tempted to get some of that gskill 3800 CL14 lol
Would running MCLK/FCLK at some other, simple-ish fraction help in non 1/1 cases over picking some random number, say 5/4 or 6/5 (in your example 1760 or 1833MHz)?
@Actually Hardcore Overclocking - was it just a simplification, or does the memory layout really mix a daisy chain and t-shape layout for each channel? If so, wouldn't it have some negative impact on stability and performance?
daisychain or t topology does depend entirely on the motherboard, most new am4 motherboards use daisy chain though.
I know I'm just behind the times, but does any of this apply to Zen+ (2700X)? Is there already a similar, older, video specifically about that?
At what RAM clock rate would it be more beneficial to downclock and max out FCLK? Is 2666 RAM with 1900 FCLK better than 3200 with 1:1?
does this mean that using somehing like 2400 DDR4 memory kit should be viable as well? MCLK =1200 Mhz , so if UCLK would be able to run at 1200 Mhz, therefore at 1:1 ratio. According to your rule of FCLK-UCLK=>166 , you would be getting with the result of this equation somewhere in the 700, even with crappy CPU which can do FCLK 1766. Which in theory means, you should be able to negate the silicon lottery, because you just dont care if the chip is good or bad. Can someone please explain if this is possible and what effects on performance this would have?
Hi,
Have you made a video about tuning memory for the Ryzen 3000?