Ryzen Owners Zen Garden

[TheLostSwede]

It appears not. I am able to boot with 1833mhz FCLK. My memory won't do that though.

Right, it just seems like there's some setting holding you back more than anything else. What about tRFC?
It'll be quite high on Micron based DIMMs from what I've seen though, like 570-ish something. But it still doesn't explain the super high latencies.
Maybe try something stupid before you splash out on new RAM, do a hardware CMOS reset, i.e. use the board jumper/pins/button and then set everything up again to see if somehow some old settings got retained when you updated the UEFI. It can happen, even though it's rare and shouldn't happen.

Edit: Since those posts last month, I discovered that I can run my memory at XMP but only with PBO off. At XMP settings I get 3600mhz Cas 18, Best time: 108.5, Random Latency: 78.14.

That's still high, you should be lower, even with that high CAS latency. Not by much, but at least 4-5ns better.

 

[tabascosauz]

@tabascosauz I apologize if you remember this conversation, but I am going to refer you to my posts in this thread from last month. We determined that my memory does seem to be the issue. https://www.techpowerup.com/forums/threads/ryzen-owners-zen-garden.231658/page-85#post-4247124

Sorry bout that, lots of super long threads floating around on here. Aside from 1:1 (which, on a side note, if you're ever messing with a lot of different freq/timings profiles, easiest way to check it's 1:1 is RM, which has a Coupled Mode field that shows ON or OFF), only other thing I can think of is something about timings.

I was fiddling with some RAM settings after the new profiles in DRAM calc 1.7.2 came out, and in particular I can recall one instance where I tried 3200 CL14 on my DJR kit, on which I usually run 3600 CL16 and are rated for 3600 CL17. I think the timings were 14-17-14-17-32-419-1T @ 3200. AIDA results were about right on the money in terms of bandwidth, but with a latency of 72.5ns, which seemed really odd considering my E-die-masquerading-as-D-die kit does 16-16-16-36 @ 3200 for 72.0ns. I'm thinking that CL14 should at least be under 70ns, even if it's not 14-14-14 B-die, but evidently AIDA had other plans.

Which timings are you changing when you try different freq/CL combinations? I used to put all the subtimings I could from DRAM calc recommendations, but now I just leave most of them apart from the main timings (tCL, tRCD, tRP, tRAS, tRC, tRFC, command rate) on what the XMP is. And even though the new 3600/16 fast profile uses looser main timings, it's still faster than the old profile with all custom timings.

Just a thought, leaving XMP subtimings could be better suited to your particular sticks, or they might not be. Yours is like a half step below in CAS, mine have XMP at 17 (albeit sharing most timings with CL16 G.skill kits) so not too far off.

Oh make sure you have Power Down Mode off. By some accounts it can really wreak havoc on latency in short memory benchmarks. Gigabyte hides it in the AMD Overclocking menu instead of beside Gear Down Mode in the regular memory, but Asus may be different.

Right, it just seems like there's some setting holding you back more than anything else. What about tRFC?
It'll be quite high on Micron based DIMMs from what I've seen though, like 570-ish something. But it still doesn't explain the super high latencies.
Maybe try something stupid before you splash out on new RAM, do a hardware CMOS reset, i.e. use the board jumper/pins/button and then set everything up again to see if somehow some old settings got retained when you updated the UEFI. It can happen, even though it's rare and shouldn't happen.

That's still high, you should be lower, even with that high CAS latency. Not by much, but at least 4-5ns better.

Man, these threads with latency issues lately have confused the hell out of me. Only common denominator seems to be that they're Corsair sticks. 80ns+ on any 3000 CPU while at 1:1 3200/16 or equivalent shouldn't be possible.

I feel like Rev.E might be a bit more freq-focused than our CJR and DJR are with their high tRFC, and seem to be the perfect candidate for frequency overclocking. But I'm starting to think that maybe some of the timings don't like one another.

 

[AsRock]

I only picked up the Patriot 3200 2x16GB set but they even run Cas 16 1800\3600@1.35v ( Hynix CJR), so yeah if i was in the market for faster ram i would be surly concidering the Patriot ram right now.

 

[Nordic]

Right, it just seems like there's some setting holding you back more than anything else. What about tRFC?
It'll be quite high on Micron based DIMMs from what I've seen though, like 570-ish something. But it still doesn't explain the super high latencies.

Maybe try something stupid before you splash out on new RAM, do a hardware CMOS reset, i.e. use the board jumper/pins/button and then set everything up again to see if somehow some old settings got retained when you updated the UEFI. It can happen, even though it's rare and shouldn't happen.

That's still high, you should be lower, even with that high CAS latency. Not by much, but at least 4-5ns better.

I reset the settings with the jumpers on the motherboard nearly every time I try to mess with memory settings. I have tried on two different motherboards, both getting similar latency. I set tRFC to what the Ryzen Dram Calc tells me to set it to.

Aside from 1:1 (which, on a side note, if you're ever messing with a lot of different freq/timings profiles, easiest way to check it's 1:1 is RM, which has a Coupled Mode field that shows ON or OFF), only other thing I can think of is something about timings.

Except for the one time I tested how high I could push FCLK in uncoupled, I have ran FCLK coupled every time.

Which timings are you changing when you try different freq/CL combinations? I used to put all the subtimings I could from DRAM calc recommendations, but now I just leave most of them apart from the main timings (tCL, tRCD, tRP, tRAS, tRC, tRFC, command rate) on what the XMP is. And even though the new 3600/16 fast profile uses looser main timings, it's still faster than the old profile with all custom timings.

If I can find the setting in my bios, I set it to what Ryzen dram calculator tells me to set it to.

Oh make sure you have Power Down Mode off. By some accounts it can really wreak havoc on latency in short memory benchmarks. Gigabyte hides it in the AMD Overclocking menu instead of beside Gear Down Mode in the regular memory, but Asus may be different.

I will try this.
EDIT: I am running default bios except for XMP enabled and PBO disabled. This has been stable for weeks. I only changed power down mode from auto to disabled and it failed to post.

Man, these threads with latency issues lately have confused the hell out of me. Only common denominator seems to be that they're Corsair sticks. 80ns+ on any 3000 CPU while at 1:1 3200/16 or equivalent shouldn't be possible.

Could it be that I am running 4 memory sticks instead of two?

 

[tabascosauz]

@Nordic it really shouldn't affect your latency by that much on its own. Surprisingly, there is very little documentation on the latencies of 4 sticks vs 2, but I've never heard of a latency penalty for 4 sticks before. It is true that the X470 Taichi is a better board for 4 sticks because of T-topology, but the topology debate strictly only affects raw frequency.

But if the XMP profile is complete ass and Corsair really just inputted random large-ass numbers into the timings, I can see a couple nanoseconds coming out of that. Questionable 3200/16 kits with timings like 16-18-18-38 already can come in at up to 75-76ns, so add even worse-optimized subtimings, and 1-2ns for being a 3900X, and you end up in the 80s.

 

[Nordic]

Is it normal to not be able to boot with Power Down mode disabled?

 

[Caring1]

Is it normal to not be able to boot with Power Down mode disabled?

My Dram Calc settings show power down as disabled.

 

[Athlonite]

Is it normal to not be able to boot with Power Down mode disabled?

Yup well atleast mine don't use it

 

[tabascosauz]

Is it normal to not be able to boot with Power Down mode disabled?

Uh, no? Never has stopped my system from POSTing before. It just prevents the DDR4 from stepping down to a lower power level in idle. I would usually think the opposite to be true, that Power Down enabled creates instability.

Is that the only thing you changed in that bootup, or were there other memory settings changed? My system usually resets its own CMOS after 3 failed memory training attempts, but there are certain subtimings that it just cannot do, and it'll end up stuck at the logo screen or on a black screen so I'll have to bridge the jumper manually.

 

[Nordic]

Uh, no? Never has stopped my system from POSTing before. It just prevents the DDR4 from stepping down to a lower power level in idle. I would usually think the opposite to be true, that Power Down enabled creates instability.

Is that the only thing you changed in that bootup, or were there other memory settings changed? My system usually resets its own CMOS after 3 failed memory training attempts, but there are certain subtimings that it just cannot do, and it'll end up stuck at the logo screen or on a black screen so I'll have to bridge the jumper manually.

It won't even post if I change nothing else but power down mode.

 

[tabascosauz]

It won't even post if I change nothing else but power down mode.

That sucks. I don't know what else to say, man. I don't think Power Down Mode preventing POST is a common occurrence, though you're probably right, that might honestly just reflect on the piss poor quality of the ICs. You're on 1407 BIOS, yeah?

The issue here being that Power Down Mode, in the worst case scenario, can be responsible for up to around 5ns of latency on its own. Disable PDM and bam, 5ns gone in tests, just like that. I'm guessing you really are going to be looking at one of those Patriot or G.skill kits.

 

[dyonoctis]

I know that the boost issue of zen 2 is supposed to be fixed since 1.0.0.3 ABBA, so I'm wondering if I'm just unfortunate with the silicone lottery. My 1700x could boost to 3.9 Ghz (xfr 100 mhz above ref) in single core. Meanwhile my 3700x is @ 4.3Ghz at best, so no xfr there. My all core boost is at 4Ghz.

 

[TheLostSwede]

I reset the settings with the jumpers on the motherboard nearly every time I try to mess with memory settings. I have tried on two different motherboards, both getting similar latency. I set tRFC to what the Ryzen Dram Calc tells me to set it to.

Could it be that I am running 4 memory sticks instead of two?

Wow, you really seem to have some turd memory, as you've clearly done what you can do. Asus seems like to name things very differently, but if you've covered the basic settings, it should still be fine.
I gained 0.9ns latency at first, going from two to four modules, but this improved to 0.7ns with later AGESA/UEFI updates, so that's not it.

I know that the boost issue of zen 2 is supposed to be fixed since 1.0.0.3 ABBA, so I'm wondering if I'm just unfortunate with the silicone lottery. My 1700x could boost to 3.9 Ghz (xfr 100 mhz above ref) in single core. Meanwhile my 3700x is @ 4.3Ghz at best, so no xfr there. My all core boost is at 4Ghz.

It's possible that your motherboard holding you back, as it's also a matter of how much power the CPU can draw.
Still, you have six CPU power phases on that board, so it shouldn't be the issue. Could also be the AGES/UEFI implementation that limits things, since technically the 300-series of chipsets, outside of the X370 never had AMD's blessing.
I had similar issues with early AGESA/UEFI implementations even on my X570 board, with peak boost being 100-150MHz below the up to boost speed. Now I boost 50-75MHz above that, but it took quite some time to get there, in terms of AGESA/UEFI updates.
I presume you're on UEFI 5406 with AGESA 1.0.0.4 B?

 

[Bronan]

Could it be that I am running 4 memory sticks instead of two?

Yes it does in my long time hunting for more memory in my systems i have had hundreds of sets with 4 banks memory failing todo higher clocks and all ended up having timings issues or won't even boot at all.
Overclocking wise its better to buy a set of 2 sticks instead of 4 if you want to get them to run at higher speeds, by that i mean even the XMP profiles fail to run them properly at the advertised speeds.
Hell many hundreds of the sets did not boot at all or could not even run stable on XMP setting at all
I actually gave up after Corsair admitted that they never had so much issues, the weirdest of it all is that i had that on EVERY top brand motherboard available on the market.
The only speeds i got working without issues was 2133 and 2400 even with sets which was advertised to run at 3200 or much higher.
So my advise switch to 2 bank set and be released to finally run them error free without insane problems.
BTW never believe the shops selling them, because they allways say that it can be done.
The only sets which did run well are the small 1 and 2 Gb ones which are absolutely way too small for usage on a modern system.
So even though there are many new faster products released you often should stick with 2 banks unless your willing to run them at 2133 or 2400.
They 4 sets suffer all from latency problems if ran at higher speeds.
If you look at the top overclockers you see them often use single 512 mb or 1 gb sticks to get the highest possible overclocks for the same reasons.

 

[TheLostSwede]

Yes it does in my long time hunting for more memory in my systems i have had hundreds of sets with 4 banks memory failing todo higher clocks and all ended up having timings issues or won't even boot at all.
Overclocking wise its better to buy a set of 2 sticks instead of 4 if you want to get them to run at higher speeds, by that i mean even the XMP profiles fail to run them properly at the advertised speeds.
Hell many hundreds of the sets did not boot at all or could not even run stable on XMP setting at all
I actually gave up after Corsair admitted that they never had so much issues, the weirdest of it all is that i had that on EVERY top brand motherboard available on the market.
The only speeds i got working without issues was 2133 and 2400 even with sets which was advertised to run at 3200 or much higher.
So my advise switch to 2 bank set and be released to finally run them error free without insane problems.
BTW never believe the shops selling them, because they allways say that it can be done.
The only sets which did run well are the small 1 and 2 Gb ones which are absolutely way too small for usage on a modern system.
So even though there are many new faster products released you often should stick with 2 banks unless your willing to run them at 2133 or 2400.
They 4 sets suffer all from latency problems if ran at higher speeds.
If you look at the top overclockers you see them often use single 512 mb or 1 gb sticks to get the highest possible overclocks for the same reasons.

On Intel or AMD? As don't assume what is the case on Intel systems, is the case on Ryzen, even less so Ryzen 3000.
I suggest you read this article, if nothing else. I will admit I was very sceptical, but my own setup shows that four DIMMs have no noticeable downsides whatsoever.

Ryzen Above: Best Memory Settings for AMD's 3000 CPUs, Tested

Is Ryzen 3000 optimized for DDR4-3200? What if we want more? We examine everything from frequency to rank count to nail down optimal settings.

www.tomshardware.com

I run 4x 8GB at 3800MHz with decent timings. 100% stable for months, no matter what I have thrown at it.

AMD Ryzen 7 3800X @ 4199.02 MHz - CPU-Z VALIDATOR

[33u9si] Validated Dump by OVERLORD-MK-MXV (2019-11-20 09:33:26) - MB: Gigabyte X570 AORUS MASTER - RAM: 32768 MB

valid.x86.fr

And dude, catch up with the times will you. No overclocker users those kind of tiny modules any more, as they're in fact not even available.
You're commenting on something using outdated information that doesn't help anyone.

Micron Memory Sets New DDR4 Overclocking World Record

Ballistix, Micron's global brand of high-performance gaming memory, has set a new overclocking world record for the fastest DDR4 memory frequency at 6024 MT/s. Leveraging performance-tuned Micron die and the innovation behind the new Ballistix Elite 4000 memory, the ASUS motherboard R&D team set...

www.techpowerup.com

G.SKILL DDR4 Memory Achieves DDR4-5886 and 23 Overclocking Records

G.SKILL International Enterprise Co., Ltd., the world's leading manufacturer of extreme performance memory and gaming peripherals, is excited to announce that 23 overclocking records in various benchmark categories were broken during the Computex 2019 time frame, including the world record for...

www.techpowerup.com

MSI Announces DDR4 Memory Overclocking World Record @ 5902 MHz on MPG Z390I GAMING EDGE AC Motherboard

MSI today announced that their in-house overclocker Kovan Yang has broken the world overclocking record for DDR4 memory. The record-breaking feat was achieved with a pair of HyperX Predator DDR4 memory modules paired with MSI's MPG Z390I GAMING EDGE AC motherboard, with a little help from liquid...

www.techpowerup.com

 

[Bronan]

lol
Those are all inhouse overclockers having access to toys which never come available for normal buyers at all.
I have just installed a new set for a friend with a x570 motherboard and again had issues, however after consulting Corsair we got a new set send to his address which actually works.

 

[TheLostSwede]

lol
Those are all inhouse overclockers having access to toys which never come available for normal buyers at all.
I have just installed a new set for a friend with a x570 motherboard and again had issues, however after consulting Corsair we got a new set send to his address which actually works.

So stop buying Corsair RAM. LPX modules have time and time again been proven not to work well with Ryzen.

And with regards to overclockers, I know a lot of the "pro" guys and yes, they have the chance to test a lot of kit, but seriously, most of it is just off a production line, just like what you and I buy. They might have a chance to test 200 kits or 200 CPUs and pick the best of the best out of that, but technically we have a chance to get as good hardware, although we're less likely to win the silicon lottery, since we can't afford to buy 200 units and test them until we find the very best parts.

What we most likely can't do, is mods like this. I know my soldering skills aren't up to it, but then again, this guy had one of the engineers do it for him...
It allowed him to hit the XMP 5000MHz settings of his RAM kit, due to the USB ports going flaky. That's an Intel Z490 board btw, if anyone was wondering.

 

[TheoneandonlyMrK]

So stop buying Corsair RAM. LPX modules have time and time again been proven not to work well with Ryzen.

Funny thing is I have used mostly vengeance pro RGB because I have to after trying gigabyte RGB auris and seeing the RGB nightmare rainbow for weeks before one stick died.
My rigs big on Corsair RGB so I'm stuck with their memory (Corsair) but because I have used four sets now it is literally the last manufacturer I would recommend.
On a fair few customer PC I went with patriot and it's way way easier to get it working without issues.
Hence why I recommend it.

 

[heky]

lol
Those are all inhouse overclockers having access to toys which never come available for normal buyers at all.
I have just installed a new set for a friend with a x570 motherboard and again had issues, however after consulting Corsair we got a new set send to his address which actually works.

LOL, where are you getting this BS from? I have built at least 10 Ryzen 3000 rigs with 4x8Gb dimms and had ZERO issues running at least 3600MT/s. Mostly MSI motherboards, but a couple ASUS too.
And here is mine, jus for the lulz:



And that is not even trying, at 1.4v

 

[tabascosauz]

For everyone who misses having a program that lists all your timing information in one place, since Ryzen Timing Checker only supports Summit and Pinnacle CPU families, there's this lightweight utility that essentially continues in that role called ZenTimings. Github for 1.0.4: https://github.com/irusanov/ZenTimings/releases

Saves you the trouble of digging through Ryzen Master or DRAM calc's cluttered interfaces just to find your timings. What it looks like:

 

[Nordic]

I noticed you are all using Aida64 to test memory latency. I have been using Ryzen Dram Calculator. I just updated from 1.6.x to 1.7.3 and my memory latency is much improved. Maybe the problem all along is what I am using to bench. The following is from Ryzen Dram Calculator 1.7.3 at my XMP speeds.

I am going to download Aida64 and bench with that too.

EDIT:

 

[tabascosauz]

I noticed you are all using Aida64 to test memory latency. I have been using Ryzen Dram Calculator. I just updated from 1.6.x to 1.7.3 and my memory latency is much improved. Maybe the problem all along is what I am using to bench. The following is from Ryzen Dram Calculator 1.7.3 at my XMP speeds.

I am going to download Aida64 and bench with that too.

That's quite an old version you had there. The custom and random numbers in 1.7.3 look ballpark alright for what I'm guessing is CL16? RDC's membench is a headache when it comes to comparing to other people's results, because it allows you to specify the test size which throws comparisons out the window, especially with single-chiplet SKUs having half the write speed. But it works perfectly fine for testing one's own computer.

 

[Nordic]

That's quite an old version you had there. The custom and random numbers in 1.7.3 look ballpark alright for what I'm guessing is CL16? RDC's membench is a headache when it comes to comparing to other people's results, because it allows you to specify the test size which throws comparisons out the window. But it works perfectly fine for testing one's own computer.

Check my edit. I added the AIDA benchmark. That is my XMP settings for 3600mhz Cas18.

 

[TheLostSwede]

Check my edit. I added the AIDA benchmark. That is my XMP settings for 3600mhz Cas18.

For CAS 18 that looks fine.

 

[tabascosauz]

Check my edit. I added the AIDA benchmark. That is my XMP settings for 3600mhz Cas18.

That's right on the money for 3200/16 and 3600/18. I'm not sure that upgrade would be worth it after all, depending on how expensive the kits are.