IF 1800+3600CL15 or IF 1900+3800CL1x ?

[Fry178]

Mainly for gaming gains, as i dont worry about encoding taking "2 mins" longer.

Finally got a 5800x that was able to do 1900 on IF/Ram, and the fact that i started using the ram with a 3800x,
i had to lower clocks, so i tweaked timings (b-die).

Running games at QHD/UHD depending on game, and if i play on or offline,
and just cant make up my mind if i should go for lower IF/ram clocks but lower timings,
or higher if but take the hit on timings.

I havent tweaked beyond cl15 (1800) nor messed with timings when running IF at 1900,
so i might still be able to improve, especially since im only doing 1.38v on ram.

 

[freeagent]

If it were me I would see how far 14-14-14 would go, or 14-15-15

 

[dont whant to set it"']

I am running 4 sticks@ 3600C14 at 1.395V dimm (B-die's)for more than a couple of weeks now, had them at C13 which proved unstable (some more tweaking was required to get them that low).

 

[Fry178]

So the lower timings are still better (perf wise), than the gains from increase in bus/if clocks?


might try messing with timings this WE.
just dont like spending hours to run HCI +400% and see it fail towards the end

 

[freeagent]

So the lower timings are still better (perf wise), than the gains from increase in bus/if clocks?


might try messing with timings this WE.
just dont like spending hours to run HCI +400% and see it fail towards the end

No if and mclk in the end are more important.. really not much diff between 1900 14-14-14 vs 16-16-16 on my 3600.

 

[Zach_01]

There is no way around the time a user must spend to tune (and stabilize) DRAM. It is what it is... and it’s pain in the @ss. If you don’t want to spend hours upon benchmarking and stress testing memory you can just use your system on your daily tasks for a few days, every time you change anything and wait for the worst. Its a long process no matter how you see it.

Ryzens (ZEN2/3) can gain some performance from MCLK/UCLK/FCLK speeds and from some timings. The speed vs timings matter has no clear answer. IMHO it should be balanced. 1800MHz CL16 is a nice place to start. And if you manage to do 1900MHz CL16-16-16-16-34-50 it would be great.
Focus on the primary timings like the 6 above and tRRDS/L, tFAW and tRFC. Others won’t make much of a difference.
tFAW should be X4 the tRRDS
tRFC should be something between x5-7 the tRC (B-Dies only). You can add a +8 on top of the multi result, for example 50x7 = 350 +8 = 358.

On B-dies I would go straight to 1.45V and work the way up to 1.5V only if needed to achieve my target and if I had additional active Ram cooling.

Depending the B-dies quality you can go all the way up to 1900MHz and CL14-14-14-14-30-44 and 220-228 tRFC without stupid amounts of DRAM voltage. <= 1.5V

GearDownMode (GDM) is better to be enabled for better stability and avoidance of extreme voltages.

 

[Fry178]

yeah, thought so but wanted to make sure.
gonna aim for 3800/1900@cl15-16, as i hope to do that at 1.45 maybe 1.4, if im lucky.

 

[tabascosauz]

yeah, thought so but wanted to make sure.
gonna aim for 3800/1900@cl15-16, as i hope to do that at 1.45 maybe 1.4, if im lucky.

Quite frankly I wouldn't even consider CL15 unless it was at a higher speed (ie. 3933CL15 or 4000CL15) than CL16. At the same speed, it isn't worth it unless you spend hours of every day staring at your AIDA results. You have to remember that to go to odd CL you are sacrificing GDM; GDM performance is pretty close to 1T performance, but 2T is a fair step down from either 1T or GDM. There's a reason why odd tCL is unpopular with Ryzen, the benefit just isn't there due to GDM. Unless you're willing to push even more volts than CL15 already needs, just to get 1T at odd tCL - which, judging from what you've said about VDIMM, I doubt you're willing to do.

If you can hit 3800 16-16-16 at 4-6-16/4-12-12 or 4-4-16/4-10-10 with a reasonable tRFC (~140-150ns), just stick with it. You're already getting a ~9ns latency reduction from the new CPU.

You can make an argument for 3800CL14 on the other hand, but that's hard to do at 1.5V for most affordable kits, and real world performance benefit is still debatable. So 3800CL15? Nah man.

 

[thesmokingman]

Adding to the above, no point in buying CL15 as you can't run it w/o disabling GDM and running GDM is a pain in the ass anyways.

 

[Fry178]

@tabascosauz
ahh, forgot about that.

@thesmokingman
well at least for 3600, 1T runs fine with gdm disabled.

 

[Zach_01]

@tabascosauz
ahh, forgot about that.

@thesmokingman
well at least for 3600, 1T runs fine with gdm disabled.

Thats why I didnt mention at all about CL15 and only talked about 14/16. Didnt clarify it though.
Yeah, one of the b-die abilities is to keep GDM disabled even at high speeds and tight timings. One other is the voltage. The 1.45V is like others having 1.35V. They will warm up but nothing that cant be controlled with a some more air flow. Things are getting more complicated after 1.5V all the way up to 2.0V. But we are not after WorldRecord, are we? So the up to 1.45~1.47V max is rather simple and common for a nice tune.

Have you run any AIDA64 before, just for reference? Its not the definitive goal, as @tabascosauz already said, but just to know if you're on the ++ side, performance wise.

 

[Fry178]

@Zach_01
Oh, so b-die is one of the exceptions to the rule.
and was like, wtf are they talking about, gdm off runs just fine

just went for 15 as it was the lowest i could get it to boot (with 1.4v),
and was still using 3800XT...

just to make sure, i can completely disregard WR/CWL/RP
tuning?
And 408 for rfc seems a bit high, is the calc just recommended or a number i shouldn't go below?
RRDS/L i get (4/4 or 4/6), but whats the 3rd timing after?
board has no proper order for sub timings.


Thanks everyone for chiming in, will probably start testing stability this weekend..

 

[Zach_01]

Mind sharing a screenshot?

 

[Fry178]

sure.
this was 100% hci stable (dos boot), didnt have time to run longer.
funny tho as it seems vram is less fluctuating with higher clocks, its closer to whats set in bios, but soc/vddp/g are a tick lower than whats set (1.05/1.0/0.95).

 

[tabascosauz]

Mind sharing a screenshot?

sure.
this was 100% hci stable (dos boot), didnt have time to run longer.

It makes me wonder if the reason the two of you are able to run GDM off is because tRFC is so loose. Tighter tRFC requires progressively more VDIMM, so it stands to reason that with loose enough tRFC, VDIMM requirements could be so low that 1T is feasible.

Would be interesting if it still holds true at 150ns (tRFC ~270 and 285 respectively for the two of you). I wonder if loose tRFC with GDM off, or tight tRFC with GDM is better. I only know that tRFC makes a massive difference.

@Fry178 you're already halfway there with the RRDS/RRDL, may as well punch in an easy 4/12/12 for WTRS/WTRL/WR. Can probably do 4/10/12 or 4/10/10 but I'm not sure if you'll need more VDIMM for that.

Also, use the latest 1.2.2 ver of Zentimings.

 

[Fry178]

i had trouble before when i was doing less than 300, but now that i have a baseline thats decently stable, its easier to fine tune, and i will try lower rfc.

thanks for wtrs etc, but still not sure what the 3rd timing is, from that rrds/l group the others recommended, as the Ultra has wtrs/l as next entries and they dont go to 16 as per recommended 4-4-16..

 

[tabascosauz]

i had trouble before when i was doing less than 300, but now that i have a baseline thats decently stable, its easier to fine tune, and i will try lower rfc.

thanks for wtrs etc, but still not sure what the 3rd timing is, from that rrds/l group the others recommended, as the Ultra has wtrs/l as next entries and they dont go to 16 as per recommended 4-4-16..

Yeah, Gigabyte's timing page is a bit scrambled. Asus is much easier to navigate in that regard, it lists them the same way as Zentimings does.

I just think of tRRDS/RRDL/FAW as a group, and tWTRS/WTRL/WR as a group. You've already got the former down pat, it's just the latter that could use some work. Do 4/12/12 if you want to play it safe, or you can play with the tWTRL and tWR if you want down to about 4/10/10.

Could also do tRDRD_SCL and tWRWR_SCL at 4 and 4 pretty easily, down to 3 if you're feeling ambitious.

 

[Fry178]

yeah, was missing the faw..
kool, thanks, will try.

 

[oobymach]

Are your rams b-die or something else? You can use thaiphoon burner summary page to determine what you have.

You can use 1usmus ram calculator to help fine tune your ram.

DRAM Calculator for Ryzen (v1.7.3) Download

DRAM Calculator for Ryzen helps with overclocking your memory on the AMD Ryzen platform. It suggests stable memory timing sets optimized for your m

www.techpowerup.com

If tuning trfc use the third tab and you only need the first value, set that for all 3 and test.

 

[Fry178]

virtually all viper 4000 kits are vbdie or at least similar quality micron/other, or it probably wouldn't even do 3600/cl15/1t@1.35.

you might wanna get a little deeper into ram tuning before recommending certain stuff, as ram calculator is more or less useless when it comes to secondary timings, especially for clocks above 3600, outside that it hasnt been update for a while now.

heck, it even said 4000 isnt possible, when both profiles (xmp and extreme) are 1000% hci stable (stock timings) @1.35v

 

[John Naylor]

Regardless of platform, the usage / applications dictate whether you will benefit from higher speeds or higher timings ... for overall usage, I look at the ole

CAS x 1000 / Speed

15 x 1000 / 3600 = 4.17

For 3800 that would be about CAS 16. At that point can compare proces of the two options and see what's more attractive ... then see how far folks have been able to go with both alternatives.

Again, it's meaningless number in and of itself for other than compariso purposes, but is a good rule of them .... better to 1st consult your primary application vendor and receive input.

But while that works pretty well on Intel, as everyone knows, it's a bit difference on AMD platforms. Applications / Usage also dicatttes platform choice and don't wind up doing many AMD builds here....in fact only 1 AMD build (0 Intel) in last year, only upgrades. However, from the bit that we have fiddled with, while gains were seen in benchmarks, weren't seeing any real increases in application performance... at least not enough to justify the time involved. And while being a nerd, Im inclined to do that on my own system, not gonna invest that time on someone else's.

What we tend to do is comsult RAM vendor web sites aand see what forum users have been able to achieve. The companies' techs tend to hang out there and they have a bit more knowledge on the topic than th average user.

 

[tabascosauz]

virtually all viper 4000 kits are vbdie or at least similar quality micron/other, or it probably wouldn't even do 3600/cl15/1t@1.35.

you might wanna get a little deeper into ram tuning before recommending certain stuff, as ram calculator is more or less useless when it comes to secondary timings, especially for clocks above 3600, outside that it hasnt been update for a while now.

heck, it even said 4000 isnt possible, when both profiles (xmp and extreme) are 1000% hci stable (stock timings) @1.35v

Actually scratch that, can you do 4000CL16 on your 5800X at 1.45V or less? That would be much preferable to 3800CL15 if your IF can handle it.

I'm not sure where you'll end up with tRFC but I found it much easier to run tRFC with higher speeds. It scales with MT/s (so 300 @ 3600 would be ~163ns but 300 @ 4000 would be 150ns). Although I can do 290 at 4000 (145ns) I can barely break 290 at 3600 (160ns or so) before it stops POSTing because even though it is technically slower in ns, each kit has its own tRFC wall at some point regardless of speed.

 

[oobymach]

virtually all viper 4000 kits are vbdie or at least similar quality micron/other, or it probably wouldn't even do 3600/cl15/1t@1.35.

you might wanna get a little deeper into ram tuning before recommending certain stuff, as ram calculator is more or less useless when it comes to secondary timings, especially for clocks above 3600, outside that it hasnt been update for a while now.

heck, it even said 4000 isnt possible, when both profiles (xmp and extreme) are 1000% hci stable (stock timings) @1.35v

Yeah it hasn't been updated for 5000 series cpu's, I been messing with my ram on and off since launch trying to get stability at higher speeds. That little tool has been immensely helpful for me.

 

[Fry178]

@tabascosauz
ram, yes, IF, nope. nothing over 1900 is stable, getting whea bus errors.

well rn rfc is 408, andi think below 270 was no-post (@3600cl15).


update
so with 3800cl16 and tweaked timings except rfc, im getting about 2GB/s more for read/write/copy and down to 55ns now.
and cb20 is better than what i had with 3600cl15 and pbo (bad cpu).

awesome help "guys"..
now let's see if its stable xD

 

[Fry178]

So my awesome +300$-super-duper-board decided to wipe ram changes i did (timings for 3800),
and i had to redo everything.

RFC of 285 made HCI behave funky on (dos) boot, freezing/slowing on loading tests etc.
290 worked fine, so seems to be limit.

Whats the best way to check IF/Bus stability?
Been messing with "tweaking/ocing" for years, but that was mainly cpu/gpu/ram clocks..