Samsung DDR3 G-die

[Noxinite]

Hello peoples.

I had a random stick of OEM Samsung DDR3 lying around and decided to try it out as a laugh. Funnily enough it would boot 2800C9 @2V and then I realised it was 2GB ds as well - huh. So I tried tightening timings and found that (unlike the typical d-die 4GB ds sticks) :

• RP was weaker.
• RFC was stronger.
• RDRD set at 4 is actually possible and stable. ?
• The three RDWR timings can be set tighter as well.

I bought some more 2GB ds sticks to test and even some 1GB ss sticks. The 1GB sticks clock a lot better than the 2GB sticks, however the performance is awful in 32M (as they are single sided). No noticeable improvements in tightening timings over the 2GB ds sticks either. I also did do some quick POST memory frequency testing and got +3600MHz working in dual channel with CPU on phase.

I have tried 4 sticks of the 2GB ds variety on phase so far and all have worked fine at those temps (around -40C to -35C). Unfortunately I overwrote an XP image that had a lot of my phase testing results on and have been busy with Team Cup so do not have much to show for that atm.

I'm also seemingly having issues with the Impact VII I'm using atm with high freq ram (>2880MHz), but it might just need some more patience.  And also my OCF is half-dead so atm I can't test on that board.

So here is the best ambient result I have so far, 2GB sticks first (the efficiency seems a bit off so I need to test some more):

1GB sticks (G3258, so on i7 will be ~1.5s faster):

 

And this is the only phase result I seem to have (although I'm sure I passed 2933C8 in single channel in my first round of phase testing):

Edited October 28, 2018 by Noxinite

[TaPaKaH]

As I said in the other thread: I've seen Samsung 1Gbit G-die do 1400 9-12-12 before and I've seen these chips on 2GB G.Skill 1333C9/1600C9-rated modules from 2011-2012.

So who knows, maybe somewhere in the wild there is a G-die set on proper PCB that is capable of 1466 8-12-12 with tRFC of just 88 

[yosarianilives]

Crazy to think that there may be really good ICs out there that nobody has tested because they were too busy binning d-die, psc, bbse, etc.

Well done on testing, and will be interesting to see what comes of this IC once you and other have more time to test!

[suzuki]

Better not test this properly,you will make my collection of Samsung become irrelevant :)).

Can you make 2 runs on same os ,identical settings so we know how much they loose from density, and after “ tight “ on both at 2800 mhz ?

Edited October 23, 2018 by suzuki

[Noxinite]

16 minutes ago, TaPaKaH said:

As I said in the other thread: I've seen Samsung 1Gbit G-die do 1400 9-12-12 before and I've seen these chips on 2GB G.Skill 1333C9/1600C9-rated modules from 2011-2012.

So who knows, maybe somewhere in the wild there is a G-die set on proper PCB that is capable of 1466 8-12-12 with tRFC of just 88 

I had a quick look on EBay and found some G.Skill sticks, but they were also using the Samsung PCB.

6 minutes ago, yosarianilives said:

Crazy to think that there may be really good ICs out there that nobody has tested because they were too busy binning d-die, psc, bbse, etc.

Well done on testing, and will be interesting to see what comes of this IC once you and other have more time to test!

That was part of the reason I wanted to sell some, to get others interested.

4 minutes ago, suzuki said:

Better not test this properly,you will make my collection of Samsung become irrelevant :)).

Can you make 2 runs on same os ,identical settings so we know how much they loose from density, and after “ tight “ on both at 2800 mhz ?

Sorry.

I might be able to do those runs, but I did have some issues with g-die not liking RTLs as tight as d-die at 2800C9 - as with most of the testing I need to double check.

Edited October 23, 2018 by Noxinite

[TaPaKaH]

http://hwbot.org/submission/2456555

Here's a hwbot score with what seems to be G-die. I remember spotting more, but can't find them anymore.

[SparkysAdventure]

This is some really cool stuff. I may be interested.

[Noxinite]

I've had some good progress; into the 18s with full ambient.

I had a few runs at 2948C9, but RTLs and IOLs were out of whack and timings wouldn't tighten.

Edited October 26, 2018 by Noxinite

[SparkysAdventure]

Push for cl8!

[Bullant]

Nice job Nox

[Noxinite]

Tried some G-die on phase quickly, but kept getting 'not exact in round' even when trying 2948C9, which I could almost do on air.

This run was looking to be around 18.5s:

It seems like RCD is very important for this IC as well - compare the above screen to the below:Tried some C8, but again was getting 'not exact in round' a lot - it was also looking similar to 2949C9 all tight for efficiency:

I also tried to get my OCF working again, but the USB headers as well as the back IO seem to be broken - and it doesn't like the USB 2.0 PCIE card I was using with Z170. It also instantly failed 2948C9 when I loaded it up.

Edit: also seems like the issues I was having with cold PSC (code 15) at >2800MHz is the Impact struggling with IOLs/RTLs. Sometimes I need to manually set it and sometimes it needs the 1st channel IOL to be raised to 5.

Edited October 28, 2018 by Noxinite

[speed.fastest]

This is probably same ic as my friend have, it scaled to 2.4v though : http://hwbot.org/submission/2632915_yash_yulian_maxxmem_read_bandwidth_ddr3_sdram_28511_mbytesec.

[Noxinite]

Ah, hah! Found my initial single channel testing on phase.

Not sure why I haven't managed to get similar clocks more recently on better sticks... Maybe my boot voltage binning method is awful? Haha.

[SparkysAdventure]

Can you get that speed and those timings in dual channel?

[Noxinite]

Not yet. I recently reflashed the bios of the Impact and it seemed to help with training ram, so maybe it was just a board issue.

[SparkysAdventure]

Gotcha. Good luck! Very interested in seeing if this is possible.

In terms of overall efficiency, is it better or worse than similarly clocked 2x4gb memory, given the different strengths and weaknesses with the timings?

[Noxinite]

I found some G.Skill G-die on Brainpower PCB (I think) and quick POST voltage testing seems to show they are better than all 14 of the 2GB OEM sticks I binned. XD

They needed 1.88V and 1.91V to POST 2800C9 tight.

No 32M passes yet though, idk wtf is going on.

[SparkysAdventure]

@Noxinite I believe you should seek out some Hynix 1GB BFR. From the limited information I've read it scales just about the same as other Hynix DDR3 in general, and based off of your findings here, other timings may be able to go tighter. It can be found in at least this kit here:

https://vic81324.pixnet.net/blog/post/32345479-kingston-藍色風暴t1-hyperx-ddr3-1600

Edited November 22, 2018 by AutisticChris

[Noxinite]

Well CFR is double sided and Shamino showed a long time ago it scaled nicely with cold. But it's still a lot slower than cold PSC.

As for g-die, I tried the best Samsung pcb sticks on phase quickly yesterday and managed to train ~3200C9, >2933C8, and ~2600C7. But nothing was passing, including my previous 2956C9 ambient 32M profile.  More patience needed maybe, or I could just bin them properly instead of boot voltage lol.

Also, a while back I tried g-die on my OCF but was having some weird results. Especially with the G.Skill G-die. However my Z97 OCF is soooo battered it could just be that.

Edit: seeing as there is an 1150 cold mem comp coming up, I might try g-die on LN2. Watch this space...

Edited February 9, 2019 by Noxinite

[Alex@ro]

Great findings here, way to go ! If you manage 2933 9 12 12 trdrd4 that should be very close or similar to frozen psc, when i was playing with it it showed like 1 sec difference between 5 and 4

[lanpartyut]

good job

[Guest]

Looking forward to results here, good luck!

[Noxinite]

Spent the weekend binning the sticks properly on 32M. Best two did 32M at 1.92V and 1.93V:

The G.Skill sticks have RCD weakness (needs 13 at 2800MHz) and seem to be trash (2.17V and 2.20V). However, interestingly they run proper RTLs/IOLs like D-die, and unlike the Samsung PCBs that run higher RTLs and whacky IOLs.

Stuck the best two sticks on phase with the proper copper EK heatspreaders and spent ages tweaking to get this:

But even phase PSC is quicker than this by 300-400ms.

[Noxinite]

Resurrected my Z97 OCF to try quad dimm g-die.

Seems okay, but so far clocks are lower than dual dimm.

Also, I can't run 43/44/4/3, 43/44/3/4, or 43/44/3/3. They all train, but don't reach initial. Need that 2 dimm board magic touch.

No wazza on this run:

With wazza:

Edited February 18, 2019 by Noxinite

[funsoul]

Thanks for all the great testing Nox! I'm finally getting back to the z97ocf but can't for the life of me remember the voltages for samsung (iirc...psc is +0.25/+0.25/+0.25 but my samsungs hate those voltages). Been trying for a few days and totally stuck so figured to beg for some help. At 2800 9-12-12 (on air), can get to windows about 50% of the time but the least stress (like opening an app) and insta-bsod.

As an aside...is it worth running sammy or psc cold under ss or cascade? Seems much easier than hitting it with ln2 (iirc...ln2 seems most worthwhile for psc?).

Any ideas/tips/etc?

Tia!!

-Stefan