Samsung DDR3 G-die

[SparkysAdventure]

On 10/23/2018 at 3:30 PM, 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 

Necrobumping for info, I still haven't seen anything G.Skill besides early 2013 that has G-Die in it. Also, 2800 tRP 12 seems to be a surprisingly difficult bin category.

Semi related, G-Die has been documented in 1600C8 Trident by @Apfelkuchen. He also saw a listing for 1333C7 Ripjaws that was very likely to be G-Die but seller fell silent. Both B63, by the way.

Speaking of, G-Die seems to have weird PCB sensitivity. Sticks on B63 have documented 1t weakness when on 4-dimmers, but appears to behave as good or better than non-ECC Samsung PCB when with 2-dimmers.

[redux]

On 6/12/2023 at 4:33 PM, redux said:

I for one am nearly done binning it and I'm hoping to do some proper dual channel tuning this month. Unfortunately most of my dual rank sticks are pretty shit, but we'll see what I can do.

Given it's been... six months since I said this and all of my effi testing has been postponed indefinitely, I thought I should at least share my binning results from earlier this year.

--------

For those who just want the raw data, I have a spreadsheet here. I have screens for (almost) all results here.

 

Tested on Maximus VII Impact with binned 4790K. Note 2.39V VDIMM cap for 1400MHz CL8 testing in particular.

Sample size: 148 sticks, all JEDEC 1333 9-9-9 1.5V rated. Composed of 121x SR (73x Samsung M378B2873GB0-CH9 and 48x Transcend TS128MLK64V3U) and 27x DR (14x Samsung M378B5673GB0-CH9, 13x Samsung M391B5673GB0-CH9).

Four tests were used: 1333MHz CL8, 1400MHz CL9, 1400MHz CL8, 1466MHz CL9. At each combination, tightest tRCD and tRP at "any" voltage were found first; voltage was then minimized for these tight settings. Binning recommendations based on these tests: 

• Ability to run 1333MHz 8-11-12 at any voltage (est. top 20% stick). Among sticks that can do these timings, above average sticks can do <2.15V.
• Voltage for 1400MHz 9-12-13: <2.00V is above average, <1.95V is potentially interesting, and <1.90V is strong (top 10% on this matter). This is good as an easy test, but not as a one-stop binning procedure.
• Ability to run 1400MHz 8-12-12 at any voltage (est. top 5% stick). If just looking at CL8 scaling, ability to run 8-12-13.
• Ability to run 1466MHz 9-12-14 or 9-13-13 with tRFC88. Among sticks that can do these timings, above average sticks can do <2.10V. Top tier sticks (top 2-3%?) can do 1466MHz+ 9-12-13.

1. On average, compared to Samsung SR OEM, my Transcend SR OEM sticks seemed to have worse tRFC and tRP scaling, but better tRCD scaling. There's too many factors to point to one specific reason for this -- could be manufacturer binning, PCB, just an artifact of sampling, etc.
2. My luck with Samsung DR OEM non-ECC was very bad, with only about 50% capable of 1400MHz 9-12-13 at all, and a few sticks with subtiming weaknesses.
3. Samsung ECC PCB (M391B5273DH3) does not seem desirable for OC, due to a 1400-1414MHz clockwall and need for +1 RTL versus non-ECC (but RTL binning can help).

 

For anybody interested in an even more granular breakdown:

Spoiler

Findings, broken down by SKU:

M378B2873GB0-CH9:

• 1333MHz CL8: ≈40% could do either tRCD11 or tRP12; ≈20% could do 8-11-12. Average of 2.21±0.01V. (standard error, not test error)
• 1400MHz CL9: ≈90% could do 9-12-13. Average of 2.00±0.01V.
• 1400MHz CL8: Only 51 sticks tested based on assumed CL8 scaling. Only 35% passed; <15% could do tRP12. Pass average of 2.34±0.01V.
• 1466MHz CL9: Only 48 sticks tested, mostly 1400MHz CL9 <2.05V. ≈50% could do tRFC88 or tRP13, and <15% could do tRCD12. Average of 2.12±0.01V.

TS128MLK64V3U:

• 1333MHz CL8: ≈80% could do tRCD11 and ≈50% could do tRP12; ≈30% could do 8-11-12. Average of 2.24±0.01V.
• 1400MHz CL9: >95% could do 9-12-13. Average of 2.00±0.01V.
• 1400MHz CL8†: <20% passed; no tRP12. Pass average of 2.34±0.02V.
• 1466MHz CL9: Only 40 sticks tested. ≈25% could do tRFC88 or tRCD12, and ≈10% could do tRP13. Average of 2.16±0.01V.

M378B5673GB0-CH9:

• 1333MHz CL8: ≈35% tRCD11 and ≈15% tRP12; only one 8-11-12 sample. Average of 2.27±0.028V.
• 1400MHz CL9: Only ≈55% could do 9-12-13. Four sticks had a weakness in at least one of tRFC/tWRWR/tRDRD. Average of 2.02±0.02V.
• 1400MHz CL8†: ≈15% passed -- two 8-12-13 samples at 2.33±0.02V average.
• 1466MHz CL9: Only 6 sticks tested. 3 pass, one tRCD12, no tRFC88 or tRP13. Average of 2.14±0.05V.

M391B5673GB0-CH9 (ECC):

• 1333MHz CL8: ≈85% tRCD11 and ≈30% tRP12; three 8-11-12 samples. Average of 2.20±0.02V. All sticks need at least +1 RTL versus M378-.
• 1400MHz CL9: 100% 9-12-13. Average of 2.01±0.02V. Two samples had weakness in tRFC. ≈60% needed +1 RTL versus M378-.
• 1400MHz CL8†: <25% passed -- three 8-12-13 samples at 2.33±0.03V average.
• 1466MHz CL9: 100% failed. Of the 11 sticks tested, all clockwalled at 1400-1414MHz.

† - Again, I omitted some below average samples, but did not properly remove them from the sample space in the same way as done with M378B2873GB0-CH9.

 

OVERALL SAMPLE SPACE AVERAGES:

• 1333MHz CL8 2.23±0.01V (n=145).
• 1400MHz CL9 2.00±0.01V (n=147).
• 1400MHz CL8 2.34±0.01V (n=31).
• 1466MHz CL9 2.14±0.01V (n=90).

I'd be very curious to see results at this scale for DR sticks, if possible; unfortunately I had a much harder time finding them in large quantities or for reasonable prices versus SR sticks.

Edited May 3 by redux
misused math words

[SparkysAdventure]

On 12/3/2023 at 3:42 PM, redux said:

Samsung ECC PCB (M391B5273DH3) does not seem desirable for OC, due to a 1400-1414MHz clockwall and need for +1 RTL versus non-ECC (but RTL binning can help).

Adding to this, I also have a Transcend ECC kit of G-Die, which uses a variant of the B83 PCB. These should be 8-layer, and I'm pretty sure the Samsung OEM dual-rank PCB used with G-Die is, too. 

They do not have the same 1t weakness on 4-dimm boards as dual-rank B63. Unfortunately, they are still slightly RTL-weak, albeit less so than Samsung OEM. I can't say much about subtiming weaknesses because these are the only sticks of this model I've found so far.

42/44/4/4 also passed SP32M waza but forgot screen, needed <2v. 42/43/4/4 stopped training at about 2750mhz, maybe bclk up in OS is possible. I haven't tested on the Impact due to weird efficiency issues I haven't solved. Also, I still need to test whether they can run >2800mhz.

(12/23/2023) I should have another 3 pairs on the way, so maybe some quad rank or mild binning will be possible. Also, I think that's a large enough sample size to check for frequency ability.

(6/10/2024) Binning for 43/4 in the outer channel seems to be needed but more than half of my sticks can do it. I think it's safe to say these are the best sticks to search for if you don't want to PCB swap.

Edited June 25 by SparkysAdventure

[somerandomtechyboi]

 

some gdies with an i7 930 + x58a ud3r running 3136 ddr3 on air but havent validated due to no working wifi card since my pci wifi card wont work for whatever reason

Gdie seems to be quite stressful on the imc as ive booted 3200 in bios before with my best psc stick at only 1.7v vtt rather than the 1.775v this 3136 or 1.875v 3184 (yet to be windows booted and screenshotted cause im shooting for 3200) though the psc stick is a whole lot worse than these gdie sticks cause its just a green pcb transcend 1333 stick so itll only windows boot 3000 and itll crash as soon as i run p95 or on its own where the gdies will only error out

Mainly looking at hynix 1/2gbit bfr/dfr to see if theyll be abit less stressful on the imc otherwise ill have to chill the pscs with a tec if i wanna hit max freq, but that wont be as fun as obliterating literally all the hwbot nehalem records set with ln2 while only using pure air cooling

 

hint to how i got this speed lies in the cpuz pic and the gigabyte board

And this cpu isnt exactly the best chip i can get since it seems to be somewhat above average but not one of those insane samples thatd do 4800 uncore at 1.7v vtt so im looking for one of those prefferably in xeon form cause itd be funny to sub some cursed scores with 3000+ ecc ddr3

Edited February 1 by somerandomtechyboi

[somerandomtechyboi]

bye bye all bloomfield air and ln2 records =p

welp i was only hoping for 3200 on this chip but went 100mhz past that all the way to 3300

vtt at 1.81v cause the cpu decided that it didnt want to scale to 1.83-1.87v today

 

apparently the 2nd gdie stick was limiting me even though it was only half detected, bios is bugged so that 16x multi needs 2 sticks to boot but figured out that as soon as training finished on the first stick i take the 2nd stick out and the 2nd stick doesnt have to be fully inserted and it posts just fine, as soon as i started doing this 3200 booted with ease and i think 3232 with 202bclk boot is also possible on this particular chip

setfsb got it all the way to 3300 but going above that needs extra vtt

not a particularly good i7 930 sample either, not a golden chip thatd do 4800 uncore at 1.6-1.7v, just a mid sample that does ~4200 at 1.7v atleast thats what id suspect itd do cause its not stable pushing high uncore

so yeah a 16 year old kid with a mid i7 930 sample and a stick of gdie all completely on air has managed to obliterate all the bloomfield ambient and ln2 records

now i just have to bin more cpus and maybe a better board alongside some rams that are lighter on the imc then 3400+ may be within reach, and i also have to buy a pcie wifi card to sub this, though i may just buy a xeon and some ecc gdie sticks just to put a cursed sub just for shits and giggles cause 3300+ ecc gdie sounds downright ridicolous

 

as for the mobo theres a g1 assassin for 40$ and an ftw3 for 50$ and im looking for a board thatll do 3000+ on the 3rd channel, if these 2 wont do that kinda speed on the 3rd channel im curious what mods i can make to my current x58a ud3r(s) maybe something to do with volts?

for voltmods im gonna have to spend 10$ on a decent multimeter (ut33d+) and i wanted to do some anyways on lga 775 since i also hold the record for g31 with my g31m s2c at 542 fsb and i wanna push it higher with higher vnb vtt and gtl/reference volts, if it wont go higher i can also buy an eg31mf

Edited February 26 by somerandomtechyboi

[ground]

@somerandomtechyboi I have done the same on x58a-oc with W3680 a while ago, see https://hwbot.org/submission/4783471_ground1556_memory_frequency_ddr3_sdram_1856.5_mhz/

Ambient limit was just under 1800Mhz: https://hwbot.org/submission/4641400_ground1556_memory_frequency_ddr3_sdram_1794.6_mhz/

Have seen as high as almost 1900 on LN2:

Was able to reproduce 1856 valid with 2 different CPUs and 3 memory sticks, suspecting board is limiting me at this point.

https://valid.x86.fr/4dv07t different CPU, same (valid capable) clockwall again. The socket on my x58a-oc is in horrible condition though, so that might be to blame.

Dual channel stability on this platform is pretty easy, even above 3000. Pushing triple channel towards 3000 is very hard though.

Triple channel max valid: http://valid.x86.fr/abc7xf

triple channel max 32M:

Dual channel max 32M:

Got very close to a pass at 1550Mhz, but ran out of dice at that point sadly.

 

Edit: I also tested some ECC gdie at some point, some sticks were way better regarding voltage requirements but all hardwalled for frequency around 3200-3300, suspecting PCB.

Edited February 26 by ground

[kicsipapucs]

2 hours ago, ground said:

@somerandomtechyboi "Edit: I also tested some ECC gdie at some point, some sticks were way better regarding voltage requirements but all hardwalled for frequency around 3200-3300, suspecting PCB."

How were the aforementioned ECC modules MHz/volt before reaching their walls?

[ground]

28 minutes ago, kicsipapucs said:

How were the aforementioned ECC modules MHz/volt before reaching their walls?

the better two were doing 2800 8-11-12-13 tight at 2.3V. Other 2 2760 8-11-12-13 2.3V. My best non-ECC sticks are doing 2720 at the same settings. Edit: all 32M

Edited February 26 by ground

[somerandomtechyboi]

2 hours ago, ground said:

@somerandomtechyboi I have done the same on x58a-oc with W3680 a while ago, see https://hwbot.org/submission/4783471_ground1556_memory_frequency_ddr3_sdram_1856.5_mhz/

Ambient limit was just under 1800Mhz: https://hwbot.org/submission/4641400_ground1556_memory_frequency_ddr3_sdram_1794.6_mhz/

Have seen as high as almost 1900 on LN2:

Was able to reproduce 1856 valid with 2 different CPUs and 3 memory sticks, suspecting board is limiting me at this point.

https://valid.x86.fr/4dv07t different CPU, same (valid capable) clockwall again. The socket on my x58a-oc is in horrible condition though, so that might be to blame.

Dual channel stability on this platform is pretty easy, even above 3000. Pushing triple channel towards 3000 is very hard though.

Triple channel max valid: http://valid.x86.fr/abc7xf

triple channel max 32M:

Dual channel max 32M:

Got very close to a pass at 1550Mhz, but ran out of dice at that point sadly.

 

Edit: I also tested some ECC gdie at some point, some sticks were way better regarding voltage requirements but all hardwalled for frequency around 3200-3300, suspecting PCB.

im curious if that ratio inbetween 1:1 and 1.5:1 uncore memclk is stable or does it insta crash when using p95 largeffts? with bloomfield anything inbetween 1:1 and 2:1 will instacrash p95 largeffts 

i have managed 2800 p95 largeffts stable with 2 double sided hynix cfr which turned out to be pretty shit only doing 2800c11 at 2.1v and seems like they degrade over 2v which is not ideal when the board can push 2.6v and probs higher if i voltmod, now those 2 sticks arent stable at this setting anymore likely due to degradation so yeah lesson learned get some samsung 2gbit d dies instead

boardlimit at only 3900 is pretty sad though for a board ive been seeking out and being basically unobtainium here in indo

common boards are asus but those are trash for mem freq alongside almost every other amibios board cause they force 1.5:1 or 2:1 uncore memclk ratio, itd be nice if those boards could be bios modded for 1:1 ratio or atleast free uncore setting like the gigabyte boards as they seem to be pretty decent hardware wise, maybe amibcp? they are socketed bios so its not exactly hard to mod and unbrick the bios

as for triple channel my i7 930 tops out at somewhere around 2500 post with 3 gdie sticks, even if i push vtt to 1.7v it makes no diff compared to running at 1.5v (neccesary for aformentioned 2800c11 setting with dual rank cfrs), doesnt help that the x58a ud3rs have a shit 3rd channel that wont go past 2800, with the dual rank cfrs which i have 3 of yea 2500 doesnt even post, constantly stuck at c5 code

ecc is something ive been wanting to toy around with just because 3000+ ddr3 with ecc is just cursed asf and 4gb + 8gb are dirt cheap not to mention there are 16gb and even 32gb+ ecc sticks so high density 3000+ would be pretty interesting, do the w xeons work with 1rx4/2rx4 or are they just stuck to the usual 1rx8/2rx8 sticks? and im curious if the regular i7s can run ecc rams

 

gdie is pretty good but im also curious about any other notable 1gbit 2gbit and 4gbit ics aside from the usual stuf like mfr psc hypers etc. mainly just wanna clock the shit out of em for fun, currently eyeing hynix 1gbit bfr/dfr and 2gbit bfr/dfr as some ics that may be decent for high clocks considering theres some llano subs using these ics albiet at cl14

psc seems like its lighter on the imc compared to gdie as 2 stick 3150 runs at 1.6 or 1.7v whilst 2 stick gdie is walled to 3100 at 1.7v but all my psc sticks are trash and id rather not use subzero just because itd be funny to destroy and take over all the bloomfield subs especially subzero ones with pure air, and the best subzero i can do is probably just jerry rigging some ram cooler out of a tec though poor mans liquid nitrogen with 99% IPA and dice looks interesting too and i will toy around with proper subzero at some point that isnt just a tec northbridge cooler with a heatsink ontop of it trying to push my bios modded p5q to 600fsb (best i got was 590 iirc)

im also curious about the x/e55xx xeons cause x/e56xx xeons are stuck to 10x ram multi so i wonder if the same applies to the x/e55xx xeons and i should just avoid em and get w35xx since there are some x5570 going for 5$

atm i have about 40-50$ and i cannot make up my mind on what to spend it on cause theres a giga g1 assassin ive haggled down to 40$ and theres an ftw3 at 50$ alongside a supposedly broken classified 4x sli or whichever evga board thats just x16 pcie spam for 27$, there used to be some flaming blades available but i think ill have to wait for those to pop up again, just wanna know if those boards are any good if i wanna push high ram freq since the g1 assassin is a higher end board but the rams are further away from the cpu, evga and foxconn i dont think they have the forced uncore ratio nonsense and foxconn in particular is interesting due to 1dpc

 

ive also got an ex58 ud3r pushed to 2800 but 16x and 18x multi is broken on bios ff so i think ill crossflash it to the x58a ud3r fb bios, but if there are better bioses to crossflash to thatd be nice to know rather than having to desolder a bunch of time

[somerandomtechyboi]

19 hours ago, ground said:

Dual channel stability on this platform is pretty easy, even above 3000. Pushing triple channel towards 3000 is very hard though.

Triple channel max valid: http://valid.x86.fr/abc7xf

triple channel max 32M:

welp heres max triple channel for me at 2758 cause third channel craps out above this with 3 sticks

i discovered that the uncore needs to be atleast 3 multis above ram multi so 17x minimum for a ram multi of 14x, dunno if this will instacrash p95 but suspecting that it will instacrash p95 basically makes high freq triple channel unusable on bloomfield which sucks

does spi 32m working indicate that it doesnt instacrash if you run p95 largeffts?

ill do some more testing later cause this session got cut short due to having to go to school (woke up at 5am)

edit : forgot to add that 1:1 uncore memclk or uncore multi below 3 multis above ram multi causes constant stuck at c3 code or half posts that dont successfully post no matter how much vtt i throw at it though i will try diff bioses cause this could just be a bios bug

havent tested fc fd or fe bioses cause unavailable but ff and fh seem to completely break 1:1 uncore so it gets stuck at d8 and doesnt even attempt to train rams, and bios fb for the ex58 ud3r causes dual rank rams to not post at high freq with 1:1 uncore memclk whilst ff fixes this, so maybe this uncore multi thing is just a stupid bios bug

just gotta make up my mind if i wanna keep this ud3r or not and sell all 3 of my ud3rs in favour of the aformentioned giga evga and foxconn boards cause if i do keep it ill probably mod this thing with a socketed bios chip to make bios swapping much easier and maybe some voltmods for voltage adjustment on the fly and some bigger voltage ranges like >2.6v vdimm just to torture my garbage tinfoil f die stick 😃

Edited February 27 by somerandomtechyboi

[ground]

when I tested on flaming blade and bloodrage I couldn't get anything lower then 2:3 working on 32nm. Expecting similar behavior with 45nm walling at 1:2.

I never bothered testing for prime95 since full stability isn't relevant to me on this platform.

For me, 1:1 never worked properly, 1:1 with uncore+1 worked a bit better, 2:3 works fine and much more stable until uncore limited. Below 2:3 I couldn't stabilise 32M, 1M would run just fine, even at 3500+ (never tested on LN2 but not expecting changed behavior.

Keep in mind - until 2 1/2 generations later (Ivy bridge E) nothing intel is able to do 3200+ besides insanely binned Lynnfield and 1366. Trying to find memory that will do 3500+ at 11-15-15-31 is also quite limiting. I'd love to try in a fresh x58a-oc, the socket on mine is badly corroded and the board is generally in rough condition, given I'm the third owner who has ran it multiple times on LN2.

5000 series xeons are all locked to 10x mem multi or even less. 5500 series even has hard lock on uncore multi, can't even change it on most boards.

[somerandomtechyboi]

4 hours ago, ground said:

when I tested on flaming blade and bloodrage I couldn't get anything lower then 2:3 working on 32nm. Expecting similar behavior with 45nm walling at 1:2.

I never bothered testing for prime95 since full stability isn't relevant to me on this platform.

For me, 1:1 never worked properly, 1:1 with uncore+1 worked a bit better, 2:3 works fine and much more stable until uncore limited. Below 2:3 I couldn't stabilise 32M, 1M would run just fine, even at 3500+ (never tested on LN2 but not expecting changed behavior.

4:5 seems to work and yes this is the only decent screenshot i have even though bclk is reading 336 since 1:1 +1 also works but i have no decent screenshot of it, no real improvement over 1:1

3:2 at this speed results in the board shutting off after trying to train the 2nd ram stick and at lower speeds does post but windows insta bsods or gives 50 page fault bsod

maybe ill give bios ff or fe or whatever a retry on this board see if 1:1 +1 3:2 and 4:5 behave any differently

p95 crashes windows on both 4:5 and 1:1 + 1 uncore but 1:1 its just fine

bonus cursed 4800 ddr3 pic just because

yea gets real unstable when i push bclk and its not vioh or qpi related cause it still does this with slowmode so just uncore

meh chip so doesnt do 4800 or even 4600 uncore at 1.8v so gotta find a better chip

 

4 hours ago, ground said:

Keep in mind - until 2 1/2 generations later (Ivy bridge E) nothing intel is able to do 3200+ besides insanely binned Lynnfield and 1366. Trying to find memory that will do 3500+ at 11-15-15-31 is also quite limiting. I'd love to try in a fresh x58a-oc, the socket on mine is badly corroded and the board is generally in rough condition, given I'm the third owner who has ran it multiple times on LN2.

usually i clean boards by taking all heatsinks off + full discharge by shorting clr cmos + powerbutton and the ps_on sometimes then rinse + dunk in a bucket of water, dunno if this would help here or not cause a bad socket you just wanna swap with a rework station

also lynnfield 3000+? ive never seen one of these with a >10x multi since the i5 750/760 that i have are stuck to 10x multi, maybe the 875k? cant check cause hwbot is broken rn, im sure the imc is capable just bclk limit due to low multi

 

4 hours ago, ground said:

5000 series xeons are all locked to 10x mem multi or even less. 5500 series even has hard lock on uncore multi, can't even change it on most boards.

that sucks, but i guess if i have 5$ i can burn ill buy that x5570, probably just to trade it for a xeon w chip

[SparkysAdventure]

Mild bump, unfortunately no comparison screens on hand. Top stick inner channel, bottom stick outer. With some mild binning, you can get these B83E PCB sticks to run proper RTLs.

Out of 4 sticks tested, so far only that lighter green stick has been able to do 43/4 in the outer channel. Haven't checked if it's possible to do without bclk up in OS.

From the runs I was getting, it looks like maybe 300-500ms faster than D-Die at comparable settings. I would expect strong OS to be cracking 5.57.9xx with these settings.

Aida64 results were also much better, with read being about 500 higher and copy being about 400 higher. Latency was like 0.1 ns lower, but that was outside of variance.

Edited March 14 by SparkysAdventure

[somerandomtechyboi]

@ground

top is at 2864c10 trfc 96 2t tightened subs and p95 largeffts stable

bottom is at 2910c10 trfc 150 3t subs completely loose and crashes on p95

even with the absolutely horrid subs on the 2910c10 it still manages to beat the 2864c10 by 6 seconds and i also made a run with the same subs as 2864 and that shaved off an extra 2 seconds, i have no idea why such a small freq bump drastically improves performance like that even with the absolute garbage subtimings managing to beat half decent actually tightened ones, and i think when i was trying for 3000 i managed to get it to the 8 min 42 second mark but it errored out at the end, maybe ill try again see what happens

also found some xeon w3503 for dirt cheap and seems like theres plenty of em too so im considering just buying all of em to bin imc for maximum freq even if the low multi makes the things basically useless for anything else

just wanna make sure before buying that their ram multi isnt locked like the xeon x/e/l otherwise im just gonna buy another w3520 or even a w3565

btw where do you get tons of these chips? cause there doesnt seem to be that many around and the sellers with a sizable amount of them are overpricing the shit out of these xeons to the point that id be better off buying a w3670 or an i7 970

[somerandomtechyboi]

bought 3 w3503s, 1 was equivalent to my trash i7 930 but a little bit better (3120 1.86v), 1 hardwalled at 2900 1.7v, and the one above looks like it can just about do 3400 with a tweak or maybe even a ram swap but highest ive gotten was 3392 (didnt take a pic) as it insta shuts down or crashes when i set 3400 with setfsb, seems to like vtt at ~1.94v so looks like these chips really like high vtt so quite abit better than my best i7 930

these are really low bin and im sure most of the samples i get will be trash but for the equivalent of around 65 cents i think just mass buying these is better than buying a single w3565 or w3580 and hoping its actually a decent sample and not a pos that wont clock 3300+ on the rams, besides im just aiming for max ram freq, and ive also loosened the subs which seems to have somewhat lightened the load on the imc so ill do some retests with my usual i7 930, vdimm ranges from 2.3-2.4v though i did also try 2.44v once

and im starting to think ill run into a ram limit sooner or later with my current sticks so i wonder if 4gbit mfr or hynix 2gbit bfr/dfr would be any good for 3600+ on x58, or do i just buy some more gdies and bin them? ecc might also be an interesting option just to take cursed screenshots with 3400+ ecc ddr3 =p

ill call it a night since i wanna go sleep and ill probs continue this later, and also buy some more w3503 with my money cause i have not been able to make up my mind on what i wanna buy for awhile be it g1 assassin or buy more random boards to fix and flip, hopefully ill get a 3400 screenie later

[somerandomtechyboi]

bought 6 more w3503 and heres the 3rd one i tested doing 3400, still 3 more to test but too lazy to test right now and i wanna go sleep

think ill hit ram limit very soon since vdimm at 2.3v, also these xeons seen to really like running at 1.94v vtt, just that they degrade somewhat quickly so only good for a few runs atleast thats what i saw with that previous w3503 above this post

dunno how much farther i can bin chips and hope for higher freq cause 3500 is probably within reach but i have my doubts for 3600 atleast on air, still pretty funny that i can just blow right past all of the bloomfield ln2 ram freq records on air without even trying for the most part since this is still probably a pretty mid cpu, decent but not the very best

and i really need to buy a wifi card so i can sub these freq alongside crossflashing that ex58 ud3r so i can get the 16x multi to work, speaking of which why does the 18x multi not work on my x58a ud3r? using bios fb since the newer bioses bork 1:1 uncore memclk and make it straight up not post (stuck at code d8) and fc fd fe are gone from the gigabyte site

Edited March 30 by somerandomtechyboi

[somerandomtechyboi]

On 2/26/2024 at 3:35 PM, ground said:

@somerandomtechyboi I have done the same on x58a-oc with W3680 a while ago, see https://hwbot.org/submission/4783471_ground1556_memory_frequency_ddr3_sdram_1856.5_mhz/

Ambient limit was just under 1800Mhz: https://hwbot.org/submission/4641400_ground1556_memory_frequency_ddr3_sdram_1794.6_mhz/

Have seen as high as almost 1900 on LN2:

Was able to reproduce 1856 valid with 2 different CPUs and 3 memory sticks, suspecting board is limiting me at this point.

https://valid.x86.fr/4dv07t different CPU, same (valid capable) clockwall again. The socket on my x58a-oc is in horrible condition though, so that might be to blame.

Dual channel stability on this platform is pretty easy, even above 3000. Pushing triple channel towards 3000 is very hard though.

Triple channel max valid: http://valid.x86.fr/abc7xf

triple channel max 32M:

Dual channel max 32M:

Got very close to a pass at 1550Mhz, but ran out of dice at that point sadly.

 

Edit: I also tested some ECC gdie at some point, some sticks were way better regarding voltage requirements but all hardwalled for frequency around 3200-3300, suspecting PCB.

welp ive gone ahead and bought a w3680 untested and with some pad scratching got it working, frustratingly doesnt post >3200 even with single stick but 3450 and 3500 get stuck at c1 code and doesnt just roll d8 02/04 so im thinking im doing something wrong with my bios settings, so ive turned to screwing around with triple channel oc and here are the results

spi 32m at 2800c10, middle stick is running trfc 194 cause its a trash stick and i need to buy a better stick to replace it, i did try for 2880 and 2912 but that middle stick is really crippling my oc capabilities even when i loosened trfc, prime95 crashes within an hour but the fact that it doesnt instacrash means full stability is a possibility at both 5:4 (20x 16x) and 11:8 (22x 16x) uncore memlk, odd uncore ratios dont work for some reason and below 5:4 gets stuck at c3 upon boot for triple channel, currently testing 2752 trfc 147 (trash stick 194) cause 2800 is being annoying and refuses to fully stabilize probably due to trash middle stick

for max frequency i booted at 3008 and setfsb to 3232, need a better stick for the third channel if i wanna run higher freq as my 2nd best stick will not detect at 3008 on the third channel

this cpu is pretty average being able to do 4.8 on 4c4t at 1.5v and 4.5ghz 6c12t at 1.4v, 4200 uncore is unstable even at 1.65v vtt when i was testing 2800

i havent really pushed this thing past 1.7v vtt cause i dont think im hitting a vtt limit atm

i still find it pretty funny that my average cpu with a lowend x58 board and trash gdie sticks all on air cooling is still within firing range of your dice scores with binned cpu + sticks and an x58a oc

also found that fb is the best bios atm out of the bioses that can run 1:1 uncore memclk as fc2 and fva refuse to run the same settings for the triple stick oc, fa just freezes on the mit menu for whatever reason, havent tested ff just yet but i have my suspicions that it might also break 5:4 and 11:8

its a different x58a ud3r than usual being able to boot 230bclk no slowmode cause i found that my usual x58a ud3r is complete trash for bclk and needs slowmode to even get to 220 and refuses to boot past 207 no slowmode, but since its the first board i did my frequency records on for bloomfield ill probably just keep it for memories sake maybe even solder on a new northbridge that isnt complete trash for bclk, and yes ill have to retest my w3503s now that i found this board is better for bclk cause i was hitting a bclk wall on em when running upwards of 3400, definitely gonna bin more w3503 in search of 3500 maybe even 3600 on air but i have my doubts on 3600 unless i find a literal golden chip

oh and ill need a better aircooler cause even with proper mounting hardware i dont think this thing will cool a 6 core at 1.6v

ive also gone ahead and wasted 7$ on 2rx4 ecc d die sticks only for them to not work and d8 roll, doesnt even work on my trash p6x58de, curious if there are any consumer boards thatll run x4 sticks or if i can bios mod this thing to run x4 sticks, but oh well guess im stuck with regular x8 which is kinda boring considering 4gb 2rx4 gdie sticks exist

[SparkysAdventure]

Been working on a G-Die reference screen all day but been hitting nothing but headaches. I only managed to get this screen:

I feel like there should be another 200-300ms left in the tank, but I can't get anything passing consistently enough to say for sure. I think ambients are screwing with my sticks or IMC, as 2800c9 wasn't even close to passing Aida mem, where at my old dorm even 2900c9 was finishing consistently.

Direct booting 2800c9 42/43/4/4 is possible with these Transcend ECC sticks, not really sure why I had such issues with it previously. Will test for clock wall Soon™, 2860 already seems viable with bclk up in OS.

Edited June 25 by SparkysAdventure

[SparkysAdventure]

On 12/3/2023 at 5:02 PM, SparkysAdventure said:

Adding to this, I also have a Transcend ECC kit of G-Die, which uses a variant of the B83 PCB. These should be 8-layer, and I'm pretty sure the Samsung OEM dual-rank PCB used with G-Die is, too. 

They do not have the same 1t weakness on 4-dimm boards as dual-rank B63. Unfortunately, they are still slightly RTL-weak, albeit less so than Samsung OEM.

Interesting enough, these also have temp sensors built in. Didn't know that was ever implemented into any DDR3. Kinda curious if they'd still read correctly sub-zero.

[somerandomtechyboi]

2960c11 triple channel done, cpu at 4.6ghz

i will try for 3008 but unlikely that itll happen considering it insta errors in spi, also managed to get this config p95 stable at 2800c11 for 12 hours

cpu errors out very quickly at 3008 even with 2 sticks where my w3680 didnt error out so maybe i need to remount again? chip wise now i got a w3670 and looks to be a better chip atleast core wise as 4.5ghz all 6c12t only needs 1.344v albiet very hot running as it ran at 95c despite the voltage and aformentioned w3680 ran only mid 80s at 4.55ghz 1.424v also 6c12t

On 2/26/2024 at 3:35 PM, ground said:

The socket on my x58a-oc is in horrible condition though, so that might be to blame.

you might be dead on here cause it took 3 remounts to get it stable at 2800c11, there was a remount where it insta crashed, another remount crashed quickly, but this remount im currently running on is 12 hours stable 2800c11

cpu was just bought yesterday pads are visually in good condition albiet this is my old x58a ud3r that cant bclk for shit and i did not cover the socket so maybe thats to blame and giving it a quick water bath will get rid of the inconsistency, so no scratched pad shenanigans like on the w3680 where i have to remount it serveral times to get it posting and all channels detecting, and yes ill try remounting the w3680 again but ill buy a razor first to get the pads scratched properly so it doesnt take several remounts to get it to post

considering you are on dice and only managed 2900 with a really binned cpu im pretty sure that boards socket is holding you back considering this aircooled pile of junk with a midly binned cpu and a garbage sample entry level board is getting 2960

since the x58a oc is unobtainium probably gonna have to replace the socket with a rework station, dont think theyre that expensive to build and id totally build one if i had the money to spare so i can fix boards with chipped northbridges or swap g31 chipsets onto better pcbs to get past 542fsb

i actually ran 2912 on the w3680 a little while ago and just forgot to post it here, maybe itd also do the same 2960 but it looks to have the best chance of hitting 3008 cause it doesnt insta crash p95 if i run 3008 on 2 sticks, might have to do some more remounting though

i am considering buying another board but i dont think the ones that are available are gonna be any better (ex58 ud3r, g1 assassin, x58a ud7) except maybe the ud7 but the ex58 extreme would be some easy hwbot subs and it would be pretty cool to own one of the top of the line boards of 2008

[ground]

3700 isn’t platform limit of x58, seems like it was just the limit of my memory sticks at the time. 

Last week, Tagg Shaggy and I were building some custom Hypers and in the process also decided to take apart the best 3 sticks of DDR3L ECC gdie I had and slap the best ICs in order on new better PCBs. 

First try with best stick of the lot (2948 trp 12 binned air) - 3780. Rather disappointing improvement. Took 2nd best stick (~2900 trp 12 air) and a different IMC on LN2 and relatively easily got to ~3820 Added capacitors to the back (2x 330uF) and got to 3880. 

After that, took the best stick, added capacitors and threw it on my good IMC on LN2, and got far to close to 4000 to not try again next couple days. 

https://hwbot.org/submission/5631354

[TaPaKaH]

inb4 we find out that using Z87/Z97 for DDR3 binning was stupid and X58 is the true WR platform for DDR3 raw clocks

[TerraRaptor]

P45 to become WR holder in the end

[ground]

3 hours ago, TaPaKaH said:

inb4 we find out that using Z87/Z97 for DDR3 binning was stupid and X58 is the true WR platform for DDR3 raw clocks

give me a way to run looser timings and we talk again Believing I'm limited by trp.

[ground]

On 8/28/2024 at 9:24 AM, ground said:

3700 isn’t platform limit of x58, seems like it was just the limit of my memory sticks at the time. 

Last week, Tagg Shaggy and I were building some custom Hypers and in the process also decided to take apart the best 3 sticks of DDR3L ECC gdie I had and slap the best ICs in order on new better PCBs. 

First try with best stick of the lot (2948 trp 12 binned air) - 3780. Rather disappointing improvement. Took 2nd best stick (~2900 trp 12 air) and a different IMC on LN2 and relatively easily got to ~3820 Added capacitors to the back (2x 330uF) and got to 3880. 

After that, took the best stick, added capacitors and threw it on my good IMC on LN2, and got far to close to 4000 to not try again next couple days. 

https://hwbot.org/submission/5631354

Tried again a couple days later. Got 1992MHz at least 6 times across multiple boots, yet not a single time managed any higher. Something is walling hard again. Vdimm doesn't matter 2.3-2.6V, VTT doesn't matter 1.45-1.55V, dram vtt 0.95-1.2V. CPU temp -60 to -80 no difference either, so suspecting either board or mem to blame here.