Lets overclock some Hynix CJR

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
Ok, this is for Primary (see tag), whose CPU cannot support DDR4-3600, it is a stretch too far for some Zen+ chips.
So I sold the DDR4-3600 I had for it (Teamgroup TForce Delta RGB, Hynix CJR), which I had been running at 3200
with manual timings that gave performance close to Second (see tag) which can run DDR4-3600 and I got some
proper DDR4-3200, hoping to be able to use XMP and avoid the hassle of manual timing.

The new memory is Teamgroup Vulcan Z DDR4-3200. Ah, more Hynix CJR but I had no problem enabling XMP.
Performance using XMP is just vanilla typical DDR4-3200 performance. It wasn't quite as good as my manually tweaked
timings, which is to be expected. When I say performance, I mean performance in BENCHMARKS (more on that later).

XMP 3200
Passmark Memtest ~17400MB/s 67ns
Aida64 ~43GB/s read 46GB/s write 38GB/s copy

With my old manual timings I used with the other kit:
Manual 3200
Memtest ~19600MB/s 65ns
Aida64 ~45GB/s read 47GB/s write 39GB/s copy

With manual timings and an overclock to 3400:
Memtest ~21700MB/s 62ns
Aida64 ~48GB/s read 54GB/s write 41GB/s copy

21700/17400 = 1.247 or about a 25% increase
48/43 = 1.116 or about a 12% increase

So you might think great.
Yes, benchmarks will register the changes.
But real-world programs, sadly, didn't show a significant improvement.

Game loads, compiling software, surfing the web (which involves lots of memory)
the kinds of things you want fast memory for, simply weren't affected in any
substantial way, despite what the benchmarks were telling me.
 
Last edited:

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
I'll try to post those CJR timings, if someone else would like to see them.
They are a combination of RDC and my own work.

I went back to XMP because real-world didn't improve much, benchmarks
to me are just a way to confirm the computer is working OK and I'm not going to lose
any sleep over a few theoretical GB/s.
 
Last edited:

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
First off, let me apologize if this was an inappropriate place to post this.

It was on an MSI Gaming Motherboard, and the motherboard did its job to squeeze out
as much performance as I could even though my CPU isn't as willing. Good for MSI.

People do ask abut RAM and compatibility, and these is some information there
about some specific kits. Let's face it, mfrs can't QVL every single kit so there will be
a lot of kits that work even if they are not on the QVL. In a way, the QVL is just a basic
guideline, and never will be exhaustively complete.

Ok, I've been trying to wrap my head around something.
If I pretend to be an IMC and do the math, I get about 10 GB/s max actual for dual channel DDR4-3200.
So, how does a benchmark like Aida64 get numbers in the 40 GB/s range?

I think the answer is bank interleave. 4 banks per bank group, so the
the 2nd 3rd 4th row or column select overlaps latency with the 1st request.
(Banks more or less operate independently in that each bank has its
own set of sense amplifiers, a special buffer that holds a single row of data.)
 

citay

Active member
SERGEANT
Joined
Oct 12, 2016
Messages
818
DDR4-3200: 1600 MHz * 2 lines per clock (DDR) * 8 bytes of width * 2 modules (dual-channel) = 51,200 MB/s theoretical burst speed (51.2 GB/s).

I'm interested in what timings you arrived at for your CJR :)
 

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
Ok, I took some camera shots. Sorry for the quality.
Had to reduce the size for server to accept.

I loaded these settings from OC settings storage but was running the XMP when I did it.
So numbers on left are XMP values, numbers on the right manual.
Nice to have comparison.


IMG_20200913_164625.jpg
IMG_20200913_164641.jpg
IMG_20200913_165205.jpg
IMG_20200913_165220.jpg
IMG_20200913_165239.jpg

Everything else is on auto.
 

citay

Active member
SERGEANT
Joined
Oct 12, 2016
Messages
818
Here's the results of my RAM tuning. Note that with Intel, tRCD is just one timing, and tRP likes to be the same. Also, tRC is auto-calculated, and tRFC is a single value.

Of course i know that my Samsung B-Die have more in them (with more voltage), but i wanted to see how well they work at near stock voltages.

The system itself is in the signature. I played with four different RAM kits, each time with the minimum voltage where at least the timings noted on the package (primary XMP timings) were stable, in addition to keeping VCCSA at 1.05V and VCCIO at 0.95V.

The kits:
Kingston HyperX Fury DDR4-2666 2x8GB, CL15-17-17 (HX426C15FBK2/16), Dual-Rank
Kingston HyperX Fury DDR4-3200 2x8GB, CL18-21-21 (HX432C18FB2K2/16), Hynix JJR @ 17-19-19-39
Crucial Ballistix Sport LT DDR4-3200 2x8GB, CL16-18-18, Micron Rev. E @ 15-17-17-39
G.Skill Trident Z 2x8GB, DDR4-3600, CL16-16-16-36 (F4-3600C16D-16GTZKW), Samsung B-Die

The graph is read from left to right. Improvements are marked green, worsening red. If changes in the benchmark results were not significant, they aren't marked. The timings are in the order that my BIOS lists them. The tertiary timings are auto-optimized by the BIOS (Timing Optimization enabled).

RAM_Excel.png


Columns B to E are the four different RAM kits. You see that each benchmark improves, from 2666 to 3200 is beneficial for everything, even though i went from Dual-Rank to Single-Rank and the timings were roughly comparable (adjusted to the clocks).

Getting the timings tighter with Micron Rev. E was again a good advantage. And then another good improvement from 3200 -> 3600.

Cinebench R15 wants to see a small setback going from 2666 to 3200, contrary to other benchmarks. As can be seen later, Cinebench is 99.9% judging CPU performance, depending on what Windows does in the background. Cinebench R15 and Geekbench 3 for example are a waste of time if you want to judge RAM performance.

Looking forther in column F, i just lowered tFAW from 20 to the minimum of 16. AIDA64 Copy gets worse. This is not an outlier. Proof is column M, where i raise tFAW from 16 to 20, and the AIDA64 Copy result improves. I explained here why i came to the conclusion that tFAW 20 is actually superior to the tFAW 16 you see recommended everywhere: https://forum-en.msi.com/index.php?...-bios-power-down-control.345279/#post-1998417


In column I, i went down to tCR 2T, because i noticed the dodgy training of the tIOLs/tRTLs. CR2T has worse performance, of course, but i could compensate it with other timings, for example going from tRTP 10 to 8 was pretty beneficial (i don't have it listed seperately though).

The best benchmark for me is WinRAR (pressing ALT-B in WinRAR). It gets results extremely fast, only sometimes hampered by Windows background processes (similar to AIDA64 Memory in that regard). You sometimes have to start the benchmark two, three times until you see the value staying constant. But still, fast results, and detects every change nicely. I've also tried the 7zip benchmark, very unreliable. I just like WinRAR bench a lot.

So, those are my findings. The main benefit of Samsung B-Die when staying at this "modest" speed of DDR4-3600 is the low tRFC that is possible, as well as some other timings such as tRCD and tRP having the same capability as tCL. For something like DDR4-3600 CL15-15-15-32, i would have to raise DRAM voltage to 1.4V or higher, same for clocks above 3600. But i like that i can get these relatively tight settings at just 1.37V (put a little extra to be on the safe side), and with almost stock IO/SA voltage (1.06V/0.96V, again a little more).
 

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
Wow, pretty nice. Thanks for the information.
I don't think my CPUs can go very far to really take advantage of such nice memory.

I was curious just to see what I could do with what I consider a possible "purchase error."
The second has B-die sticks and even at 3600 XMP loose it beats the pants off the CJR,
no matter what I do with it.

I did order some more RAM (fun!) G.Skill Flare X to try. It should be B-die.
This was the memory I wanted in the first place but bad luck it was sold out
when I got the money, and the Vulcan kit was pretty cheap. Foolish.
The flare X got back in stock one day later.

Its not that the Hynix is terrible RAM, but there is just something about GSkill sticks
that makes them run better faster harder. :)
 

citay

Active member
SERGEANT
Joined
Oct 12, 2016
Messages
818
Yes, B-Die has an inherent advantage. It's just the king of DDR4. Of course it also depends on the CPU and the mainboard. But you have the best chances with B-Die.

I wouldn't invest too much money into it now though. Luckily the prices have come down a lot in the last months. The cheapest B-Die should be good enough. But getting crazy expensive kits now will look foolish soon, with DDR5 around the corner. Even the slowest DDR5 will make DDR4 B-Die look slow, i fear 😅
 

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
Oh right I'm just having some fun and I don't mind spending a little if I can learn something along the way.
I hadn't spent anything on computers in years. :)

But I totally agree with you, it looks like DDR5 is going to be amazing.
 

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
Well, the Flare X arrived and it was a huge disappointment.
The egg had run out of stock and got some new in and they appear to be Samsung C-die.
There is absolutely no room for overclocking or tightening timings with these.
You'll get 3200 XMP and that is all.

The Hynix CJR is starting to look better now.
At least these have some potential and I was able to overclock mine within the limits of my CPU.

I wanted to mention a couple of interesting things about passmark memtest.
I noticed that on the flarex pass 1 takes about 11 minutes and pass 2 takes 17 minutes or so. Weird?
Testing the CJR sticks pass 1 takes 9 minutes, and while the second pass does take a little longer, its not 17 minutes.
(Tests 1 through 9, verified on two separate computers with 3 kits of DRAM, 2 gskill, 1 team)

OK, I'm done trying ram now.
This last kit broke my heart.
:crying:
 
Last edited:

citay

Active member
SERGEANT
Joined
Oct 12, 2016
Messages
818
Yeah, any other Samsung dies don't even come close to B-Die...

Just return this last kit. CJR is way better, it should be on a level with Micron Rev. E, which is not too shabby at all.

I couldn't tell you how many hours i spent on RAM testing. After a certain point, you have to ask yourself, are you still enjoying this, or is it time to stop 😅 I could continue to clock my RAM higher and start with the timings again, but i decided to be satisfied with what i reached, and it's verified stable. This is another thing, RAM stability testing just takes too long...
 

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
It doesn't help when the previous batch was B-die and reviewers said so.
If there is any B-die left, I think it goes to the 3600 and up modules now, but who can really tell for sure.

They are being returned.

I have what I need, I just didn't know it until I tried this kit.
Funny how that works.

I was busy developing on an embedded project last couple of years
(RAM was expensive then) and as usual I am late to the party.
 

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
Here's the results of my RAM tuning. Note that with Intel, tRCD is just one timing, and tRP likes to be the same. Also, tRC is auto-calculated, and tRFC is a single value.

Of course i know that my Samsung B-Die have more in them (with more voltage), but i wanted to see how well they work at near stock voltages.

The system itself is in the signature. I played with four different RAM kits, each time with the minimum voltage where at least the timings noted on the package (primary XMP timings) were stable, in addition to keeping VCCSA at 1.05V and VCCIO at 0.95V.

The kits:
Kingston HyperX Fury DDR4-2666 2x8GB, CL15-17-17 (HX426C15FBK2/16), Dual-Rank
Kingston HyperX Fury DDR4-3200 2x8GB, CL18-21-21 (HX432C18FB2K2/16), Hynix JJR @ 17-19-19-39
Crucial Ballistix Sport LT DDR4-3200 2x8GB, CL16-18-18, Micron Rev. E @ 15-17-17-39
G.Skill Trident Z 2x8GB, DDR4-3600, CL16-16-16-36 (F4-3600C16D-16GTZKW), Samsung B-Die

The graph is read from left to right. Improvements are marked green, worsening red. If changes in the benchmark results were not significant, they aren't marked. The timings are in the order that my BIOS lists them. The tertiary timings are auto-optimized by the BIOS (Timing Optimization enabled).

View attachment 141052

Columns B to E are the four different RAM kits. You see that each benchmark improves, from 2666 to 3200 is beneficial for everything, even though i went from Dual-Rank to Single-Rank and the timings were roughly comparable (adjusted to the clocks).

Getting the timings tighter with Micron Rev. E was again a good advantage. And then another good improvement from 3200 -> 3600.

Cinebench R15 wants to see a small setback going from 2666 to 3200, contrary to other benchmarks. As can be seen later, Cinebench is 99.9% judging CPU performance, depending on what Windows does in the background. Cinebench R15 and Geekbench 3 for example are a waste of time if you want to judge RAM performance.

Looking forther in column F, i just lowered tFAW from 20 to the minimum of 16. AIDA64 Copy gets worse. This is not an outlier. Proof is column M, where i raise tFAW from 16 to 20, and the AIDA64 Copy result improves. I explained here why i came to the conclusion that tFAW 20 is actually superior to the tFAW 16 you see recommended everywhere: https://forum-en.msi.com/index.php?...-bios-power-down-control.345279/#post-1998417


In column I, i went down to tCR 2T, because i noticed the dodgy training of the tIOLs/tRTLs. CR2T has worse performance, of course, but i could compensate it with other timings, for example going from tRTP 10 to 8 was pretty beneficial (i don't have it listed seperately though).

The best benchmark for me is WinRAR (pressing ALT-B in WinRAR). It gets results extremely fast, only sometimes hampered by Windows background processes (similar to AIDA64 Memory in that regard). You sometimes have to start the benchmark two, three times until you see the value staying constant. But still, fast results, and detects every change nicely. I've also tried the 7zip benchmark, very unreliable. I just like WinRAR bench a lot.

So, those are my findings. The main benefit of Samsung B-Die when staying at this "modest" speed of DDR4-3600 is the low tRFC that is possible, as well as some other timings such as tRCD and tRP having the same capability as tCL. For something like DDR4-3600 CL15-15-15-32, i would have to raise DRAM voltage to 1.4V or higher, same for clocks above 3600. But i like that i can get these relatively tight settings at just 1.37V (put a little extra to be on the safe side), and with almost stock IO/SA voltage (1.06V/0.96V, again a little more).
I'm going to try some of these with the B-die that I do have. (Second has them.)
Thanks.
I haven't seen where I can change tREFI on AMD boards, bios wants to set them to the regulation 7.8us.
 

citay

Active member
SERGEANT
Joined
Oct 12, 2016
Messages
818
It shouldn't be too important for performance. You already get a good impact from reducing tRFC. I even stayed a bit on the safe side with tRFC 320 at DDR4-3600. It might go down to 300 or 280, that kinda region. For tRFC 320 at 3600 MHz, as it's frequency-dependent, it translates into 178 ns of refresh. So for example, 160 ns might work for you, cause it's also refreshed more often. That would be tRFC 288, tRFC2 214 and tRFC4 132. Although, you probably only have to enter tRFC and the others are auto-calculated.

IMO, this one of the biggest advantages of B-Die, how low it can go on tRFC.

More tRFC/tREFI talk here BTW, https://forum-en.msi.com/index.php?threads/ddr-training.345841/#post-2000832
 
Last edited:

flyingv2815b702e7

Active member
Joined
Aug 14, 2020
Messages
468
I read the other thread. about testing the refresh, I noticed when I was using it
that passmark memtest had something called the bit fade test and I wondered hmmm
is that to test refresh?
 

citay

Active member
SERGEANT
Joined
Oct 12, 2016
Messages
818
Yes, exactly, it tests if refresh cycle time & refresh interval are long enough & in short enough intervals to prevent bit fade.
I personally never had RAM that failed that bit fade test, maybe it's because i never max out tREFI, and try to not go right to the bottom edge of tRFC.
Speaking of refresh, do that to the other thread, i just added some more info.
 
Top