8 PACK MEMORY RANGE GROWING: SAY HELLO TO 8 PACK RIPPED EDITION & 32GB KITS!!!

[GregI]

Just out of curiosity, has anyone with the Ripped kits tried the XMP timings at 1.40v for stability? Reason I ask is the product page stating "1.40-1.50v VDIMM"

I guess this is just numbers to state voltage tolerances or something because the 3600C16 kit states "1.20-1.40v VDIMM" and I'm pretty sure 1.20v at that speed and latency is probably impossible.

 

[opethdisciple]

Just out of curiosity, has anyone with the Ripped kits tried the XMP timings at 1.40v for stability? Reason I ask is the product page stating "1.40-1.50v VDIMM"

I guess this is just numbers to state voltage tolerances or something because the 3600C16 kit states "1.20-1.40v VDIMM" and I'm pretty sure 1.20v at that speed and latency is probably impossible.

I think that might be the tested safe operating voltages for the actual chips themselves. Not necessarily the range that the clock speed will operate within.

 

[Joxeon]

Just out of curiosity, has anyone with the Ripped kits tried the XMP timings at 1.40v for stability? Reason I ask is the product page stating "1.40-1.50v VDIMM"

I guess this is just numbers to state voltage tolerances or something because the 3600C16 kit states "1.20-1.40v VDIMM" and I'm pretty sure 1.20v at that speed and latency is probably impossible.

I just go with 8 packs advice.

For 24-7 1.5v is fine on Samsung IC. Heat spreaders are also only for looks no cooling is needed.

 

[Robert896r1]

Daily fully tested (TM5/Extreme profile and large/avx2) primarily tuned for sim racing in VR

 

[Smiffy-UK]

I just go with 8 packs advice.

Agreed. I've got the team group xtreem 8 pack 4x8gb kit and I've been running them daily at 1.5v 3800cl14 for months now, don't know how different mine are compared to the ripped ones but probably similar?

 

[MrPils]

I just go with 8 packs advice.

8 Pack said: ↑
For 24-7 1.5v is fine on Samsung IC. Heat spreaders are also only for looks no cooling is needed.

As with most things said about memory by anyone, anywhere, ever... that's mostly true . Its very rare you can make a blanket statement concerning memory, XMP was created to make stable custom timings possible so if you are using XMP then the statement is 100% correct. If tuning timings however you will end up running slacker timings if you don't control memory temperatures OR you need to leave trrd's and tfaw slack to throttle throughput (if you stability test anyway, may not be required for gaming/light loading depending on case temp etc etc). There is a direct correlation between die temperature and overclocking capability when talking about B-die, it's incredibly temperature sensitive. I believe if you reported any type of heat issues or instability then the first step in a support call would be to test at XMP, therefore applying a hard TRRD throttle and bringing temps down.

B-die does not reach high temperatures compared to something like a cpu, it will rarely if ever exceed 55c no matter what you do...but its overclocking performance and latency potential is still very sensitive to operating temperatures even though its operating temperature range is much lower.

Its important to understand what was being stated by 8 Pack there - the statement says 1.5v will not harm your B-Die dimms or make them overheat. It does not state you will get the same overclocked performance if you don't cool your memory when running 1.5v. Cooling B-Die is required if you want to achieve optimum performance from as low as the 1.4v to 1.45v range depending on your case temperatures and memory loading. If you are happy leaving a TRRD/TFAW throttle in place, no cooling is required ever. That doesn't quite roll off the tongue like "1.5v is fine without cooling on B-Die" though...

Mate I ran Super PI today to compare the 2 configurations. Pretty similar results

3828c14 --- 8m 27 . 388s
4097c15 --- 8m 27 . 773s

Finally found time to do some testing myself today, I was really curious what I could achieve. Sadly it turns out the Gigabyte X570i is trash compared the MSI B450i at high memory mhz. I couldnt boot at XMP 4400mhz or even over 4333mhz no matter what I tried on the Gigabyte X570i, timings, voltages, nothing mattered. On the MSI B450i XMP works without issue and allows an overclock without touching anything except memory multiplier up to 4666mhz. At 4666mhz its then capable of booting into windows at 16-16-16-16-35-GDM for a cpuz validation - I did this in September (maybe?) last year but didn't adjust any secondary or tertiary timings or run benches, I was just going for max bootable in to windows to satisfy curiosity. I did post a link to the validation either on here or xtremesystems but I've since lost the link.

Initially 4333mhz @ 16-15-15 1T looked really promising - 85.95 seconds to complete memtest easy bench and 164.09 seconds to complete default. However when I ran MaxxMem the story became clear - performance in every metric it measured was lower than my 1:1 3733mhz settings. Latency was the killer falling back from 61.2ns to 67.3ns, as maxxmem measures its bandwidth by performing pi calcs the latency and inbalance in the Infinity Fabric and cache speeds really showed. Linpack confirmed it was a drop of around 1.5% in GFLOPS. I didn't screen cap it, I went to try reducing TRP, TRAS and TRFC expecting a bit of improvement and destroyed ntoskrnl.exe on the next reboot. Game over man - didn't even get to try GDM enabled before Windows got eaten.

Interestingly the only subtiming requiring change to boot all the way from 3733mhz up to 4333mhz was tWCL (set to match CL-2). Other than that you only need to touch slacken Primaries (+ match TRC/TRFC) to get 4333mhz, even the Tertiaries with SCLs at 2 and everything else at 1 worked fine. Proves how strong the memory controller is, shame the architecture isn't letting it spread its wings. TRCD 15 at 4333mhz is nuts.

 

[Mik3]

So the 8pack 3200 is the same bin as the 3600? It will overclock and do the same timings as the 3600?

 

[MrPils]

3200 14-14-14 is higher performance than 3600 16-16-16, but only around 0.1ns higher, therefore they are with all likelihood the same bin. The exception would be if for some reason the 3600 C16 kit used chips that were downclocked from a higher bin (they don't). Corsair do this across their upper model range to maintain their brand reputation for overclocking memory, as far as I know nobody else does.

We don't know how many bins are available (so we don't know how tightly chips are filtered between bins), or if chips within a bin have been screened for higher performance and removed before assembly. A bin gives you a low mark, not a high mark of capability...it tells you how bad the worst of the chips will be and has nothing to do with how good the best are. I suspect this is how the 3600 14-15-15 kit is made...I know 4000C15 (7.5ns) exists as a bin and 3800C14 (7.37ns) is screened from within that, but as far as I am aware 3600C14 (7.78ns) does not exist as its own bin. It would make sense if 3600C14 (7.78ns) was screened from a 4000C16 (8ns) bin via an increase of voltage, that should get a high enough pass rate to make it viable. Voltage changes will generally give +-0.25ns reliably. Memory is interesting stuff...

 

[Phil99]

I recently bought the 32GB kit (2x16GB, 3600C16) and have been playing around with tightening the timings. I've had some success so far, but am a bit confused over the lack of measurable difference between these two configs:


stable @ 1.45v


stable @ 1.50v

Everything looks so close that I wouldn't be surprised if there was not a statistically significant difference between the two.

I'm tempted to stick with the C16 setup @ 1.45v unless anyone has any ideas why the C14 config doesn't appear to be any better, or has any suggestions for further tuning?

 

[Joxeon]

I recently bought the 32GB kit (2x16GB, 3600C16) and have been playing around with tightening the timings. I've had some success so far, but am a bit confused over th

Have you mem tested?

 

[Phil99]

Have you mem tested?

Both configs are stable in memtest86+, full 4 pass test if that's what you're asking? (and daily use, i.e. Battlefield V and Anno 1800).

 

[Joxeon]

Might just be the fact then that you need to further tighten the subs to see the benefit of 14 over 16.

 

[Mik3]

3200 14-14-14 is higher performance than 3600 16-16-16, but only around 0.1ns higher, therefore they are with all likelihood the same bin. The exception would be if for some reason the 3600 C16 kit used chips that were downclocked from a higher bin (they don't). Corsair do this across their upper model range to maintain their brand reputation for overclocking memory, as far as I know nobody else does.

We don't know how many bins are available (so we don't know how tightly chips are filtered between bins), or if chips within a bin have been screened for higher performance and removed before assembly. A bin gives you a low mark, not a high mark of capability...it tells you how bad the worst of the chips will be and has nothing to do with how good the best are. I suspect this is how the 3600 14-15-15 kit is made...I know 4000C15 (7.5ns) exists as a bin and 3800C14 (7.37ns) is screened from within that, but as far as I am aware 3600C14 (7.78ns) does not exist as its own bin. It would make sense if 3600C14 (7.78ns) was screened from a 4000C16 (8ns) bin via an increase of voltage, that should get a high enough pass rate to make it viable. Voltage changes will generally give +-0.25ns reliably. Memory is interesting stuff...

Thanks, I might as well buy the 3200.

 

[Robert896r1]

16-4200 1T. Stable at 52x/48x cpu. Need a direct die to run 53/49 for daily in this heat but doable easily with direct die and loop:

vdimm only 1.53v which i'm more than comfortable running 24/7

 

[MrPils]

I recently bought the 32GB kit (2x16GB, 3600C16) and have been playing around with tightening the timings. I've had some success so far, but am a bit confused over the lack of measurable difference between these two configs:

......

Everything looks so close that I wouldn't be surprised if there was not a statistically significant difference between the two.

I'm tempted to stick with the C16 setup @ 1.45v unless anyone has any ideas why the C14 config doesn't appear to be any better, or has any suggestions for further tuning?

You still have throttling in place with trrdl 6 and twtrl 14, you need to get those down to 4 and 8 respectively. Not sure why you're running TRP high either, that will normally go at least 1 below Cas if not 2 below. tCWL can be up to 4 below Cas, though 2 below is usually optimal unless mhz limited for some reason. tRTP should optimally be 1/2 tWR, though it makes little difference.

The one thing that stands out to me is that you have your tRDWR and tWRRD set way too high - those two combined should not be higher than 15 ever. You do have to run them slightly looser when using double sided dimms but you should still be aiming to get these timings down to a combined total of around 13. Yours are currently almost double that (18 + 7), if these are at Auto in bios it indicates your motherboard is really struggling to auto train your memory at your current settings. When those two timings go way outside expected values at Auto you need to dial back whatever has caused it otherwise you end up chasing sporadic crashes or reduced performance.

 

[Phil99]

You still have throttling in place with trrdl 6 and twtrl 14, you need to get those down to 4 and 8 respectively. Not sure why you're running TRP high either, that will normally go at least 1 below Cas if not 2 below. tCWL can be up to 4 below Cas, though 2 below is usually optimal unless mhz limited for some reason. tRTP should optimally be 1/2 tWR, though it makes little difference.

The one thing that stands out to me is that you have your tRDWR and tWRRD set way too high - those two combined should not be higher than 15 ever. You do have to run them slightly looser when using double sided dimms but you should still be aiming to get these timings down to a combined total of around 13. Yours are currently almost double that (18 + 7), if these are at Auto in bios it indicates your motherboard is really struggling to auto train your memory at your current settings. When those two timings go way outside expected values at Auto you need to dial back whatever has caused it otherwise you end up chasing sporadic crashes or reduced performance.

Thanks for the response!

tRDWR and tWRRD are indeed Auto. I backed off to D.O.C.P/Auto settings and these are still 18/7, respectively. It sounds like that doesn't bode well for tuning potential based on your reply?

 

[MrPils]

Thanks for the response!

tRDWR and tWRRD are indeed Auto. I backed off to D.O.C.P/Auto settings and these are still 18/7, respectively. It sounds like that doesn't bode well for tuning potential based on your reply?

I'll be honest with you I don't have a solid formula for tuning those two values. I've read tons of forum posts, guides, anything I have been able to find for the last 12 months on Ryzen memory clocking and have yet to find anything definitive about how to tune these. I know that auto values reduce the combined total as you reduce mhz down to 2133mhz, but the values themselves seem to be as a direct result of training.

Single Sided memory sticks seem capable of running the combined total at 10, or even as low as 9 in some cases at 3800mhz. Its hinted in various guides that changing the "ratio" of these two values can affect stability, but Ive tended to find that all my sticks autotune tWRRD between 1 and 5 and tRDWR between 7 and 11 depending on subtimings, and that my final best subtimings always end up with them auto tuning to 9/1, which is the only stable setting anyway despite trying all sorts of other manual combinations... 8/1 is almost stable but never quite, other combinations totalling 9 wont post.

With my double sided sticks it wanted to auto to a combined total of 15 @ 3733mhz (11/4 or 10/5 depending on secondary timings set), with single sided it autos to a combined total of 10 (9/1 or 7/3 depending on sub timings). I have never had any luck with tuning these values either lower or in ratio to each other, I tend to find I simply cant post or clearly isn't stable.

I know that's not really an answer that helps you, but its an honest answer. My guess is that tRDWR between 10 and 12 and tWRRD between 2 and 5 is where you want to be, but it is a total guess. I have only one set of quad channel 16gb B-Die sticks and they're from 2016 and don't play so nice with Ryzen on the latest AGESA versions, so my experience is really limited with tuning double sided.

 

[Robert896r1]

rdwr = tcl - twcl + X
wrrd = tcwl + 6 + X

X in both cases is the lowest number you can make stable.

 

[MrPils]

rdwr = tcl - twcl + X
wrrd = tcwl + 6 + X

X in both cases is the lowest number you can make stable.

That doesn't seem to work out based on whats stable though... tWRRD is always less than 5 and tRDWR is always less than 12 on auto values irrespective of CL or tWCL. The combined total when they are added together always stays the same based on whatever memory multiplier I have selected. Once I have my settings dialled in it always autos to tRDWR 9 and tWRRD 1. I can then push tRDWR down to 8 for most benches, but its not daily stable. As I dial in my settings the auto values jump around between 7/3, 8/2 and 9/1 depending how far through my tune I am. If I try these other settings after my timings are dialled in I get no post.

Based on the formula you have for them, tRDWR could never be below 2, which checks out, but tWRRD could never be below 20 for CL 14 and that isn't even close in the real world using auto values, it trains itself way lower than that. There's so much confusing information around on memory settings...some of it legacy stuff people have carried forward between ddr generations, some of it specific to graphics, specific memory chips, automotive applications, CPU architectures etc etc. Most of it isn't published definitively anywhere at all, we have to figure it out the hard way.

One possibility is that we could be looking at just the additive value in bios...the "+X" in your formula. The rest of the equation could be worked out derivatively from the set timings during training, however without documentation from AMD on how AGESA handles the training procedure we are unlikely to find that out. The only evidence I have that may go against this is that I can run tWCL 10 with almost full stability without having to increase tRDWR or tWRRD, and increasing tWCL or CL does not allow me to reduce tWRRD or tRDWR.

The one common denominator I have found is that when you have a primary timing right on the edge of its capability then the auto values for these settings go nuts and start spitting out essentially random numbers. There's clearly something in training that affects what they tune to but I have no idea what.

Whatever the case, we still don't have any information on how the balance between these timings is achieved in order to tune them...the fact that the ratio between them changes depending on timings trained yet the cumulative value remains the same makes me thing its some kind of balance between read and write performance. Similar to CLK period on DDR3 where you could shift the "priority" between read and write operations (a true enough statement to get the point across quickly, but not factually correct).

I love trying to figure this stuff out, it's what got me this deep in the first place

 

[Robert896r1]

I generally don't bother with letting the board do the training outside of RTL's which I prefer the board tighten but even those can be hand tuned. The rest for bdie on intel (which is what i posted above) follows formulas and dependencies and can be hand tuned at a per timing level. It also makes it much easier on the board to post settings vs it having to figure it out. None of the timings in my screenshot are auto.

I'm not familiar with AMD To know which settings can be hand tuned vs relying on the board but mem training can vary not just from board to board but from FW to FW and from IMC to IMC so i try to avoid any such factors where possible.