I've spent a lot of time today looking at hardware. Apart from the
usual Fedora CD-R read errors, I've established:
1. ECC memory works better when it's turned on.
2. The memory you lent me (ex opetrix) hits about 1 ECC error a
second while running memtest. I'll bring a CD with me to our next
meeting.
Discussing this on the efnet #BSDcode list was interesting:
* groOgle removes 4 GB memory from obelix and burns his fingers.
<groOgle> Boy, these mammary chips run hot!
<Darius> weird
<Darius> mine are nice and cool
<kirma> yawn
<groOgle> Darius: How much memory?
<ket> CC and C3 runs hot because it's got poorer electrical qualities
<ket> duh
<ket> have to ram the signal through basically
<Darius> groOgle: 2x256Mb
<Darius> although the eflux from the CPU heatsink washes right over them
<groOgle> Darius: I suppose 8 times the memory generates about 8 times
the heat.
<Darius> heh true
<Darius> you have 2Gb DIMMs?
<groOgle> Yup.
<groOgle> I can't get the bloody machine to run with 8 GB, though.
<groOgle> And that's what I'm testing.
<ket> 2GBs run goddamn hot, they're all CC or C3
<ket> 2.6 or 2.7?
<Darius> what's a CC/C3?
* groOgle reboots tomato.
<groOgle> ket: Hold on, photo coming.
<ket> market grading on the TSOPs
<groOgle> http://www.lemis.com/grog/Photos-20050404.html#RAM-1
<ket> CC/C3 is basically 'meh, it isn't defective, hopefully.' pretty
much lowest level of testing pass, possibly not tested higher but
usually rejected
<groOgle> ket: So does the photo tell you anything.
<ket> samsung doesn't reject parts they know won't pass, doesn't test
parts they know won't pass. same for micron. all 2gb dimms are
cc/c3.
<ket> patience grasshoppa
<ket> especially since your url is 404
<groOgle> Damn.
<groOgle> The photos are there, just the main page is missing.
<ket> heh
<ket> gonna fix/ ;P
<groOgle> Hold on, any day now.
<groOgle> OK.
<ket> WTF?!
<ket> BGA?$!@&*(!@$!@?!
<ket> what. THE. *FUCK*!
<Darius> ooh sexa
<ket> NOT sexy
<Darius> looks like a routing nightmare :)
<ket> 4SD11... that's odd
<ket> in fact, that's not a valid micron part number...
<Darius> why would they be micron chips?
<ket> bottom corner has MT
<ket> micron log
<ket> er micron logo.
<Darius> OK
<ket> wintec doesn't private-label or private-source.
<Tekrad> micron has a custom labeling system....
<ket> yeah. wintec doesn't use it.
<ket> and regardless i should be able to get a hit on partial
<ket> going to have to try parametric
<ket> micron does NOT make a lot of BGA parts.
<ket> they're doublesided?
<ket> that's.. 18 per side for a total of 36 per dimm... which is
technically an invalid ecc configuration
<ket> and they'd have to be 64MB depth which micron doesn't make
<ket> not in a PC3200 BGA
* groOgle watches in amazement as memtest spews forth one ECC error
after another.
<ket> i'm not surprised.
<ket> memtest can't do ecc worth shit
<ket> but... those parts are.. well, theyr'e fucking suspicious at minimum.
<ket> the BGA layout is weird and a half. 4+5 flipped?
<ket> and with micron not making a 64MB PC3200 part... it's hard to
say if those are just outright forgeries from china or if
they're overclocked.
<ket> at most it can be a 128MB with that many, and their smallest
PC3200 BGA is a 256MB
* groOgle listens with interest.
<ket> so, most likely you're looking at overclocked memory, which is a
big no-no in ecc.
<ket> take it off SPD and try PC2700
<Darius> you should have got quality RAM like my GEIL ;)
<ket> geil is overclocked outright lying trash
<ket> well now, this is odd.
<Darius> but at least it works
<groOgle> ket: This memory has an advantage that outweighs everything
you've said so far.
<ket> groOgle - what's the voltage on it? 2.6V?
<groOgle> ket: How can I tell?
<groOgle> THe MB is set to "auto", and doesn't even tell me what that means.
<ket> oh you have gotta be shitting me...
<ket> in the bios should be a setting for dimm voltage
<groOgle> i.e. how man V.
<groOgle> Sure there is.
<ket> oh christ.
<groOgle> But there's an "auto" setting.
<ket> fuck you msi. retards.
<ket> auto means 'read spd'
<groOgle> And since I don't have any info, that's what I'm using.
<ket> ... DUH!
<ket> groOgle - memtest reads SPD
<ket> i need the SPDEEPROM output
<ket> that'll tell me the REAL part numbers _and_ voltage
* groOgle suspects that ket is saying "the setting is correct".
<groOgle> OK, how do I dump it?
<ket> when memtest starts it shows spd contents
<ket> e.g. SPEED: PC3200 (DDR400B) VOLTAGE: 2.6V
<ket> etc
<groOgle> ket: OK, I'll check when I restart. There's nothing showing
on the screen right now, but I want it to complete the
tests.
* groOgle notes approximately one ECC error per second.
<ket> HCoyote - how's your memory-fu?
<HCoyote> I can remeber that I've got 10 fingers and 10 toes ...
<HCoyote> and that today is monday
<ket> HCoyote - 36 chip 2GB, bank arrangement 4x9.. that's not valid is it?
<HCoyote> could be chipkill
<ket> it's not, checked the pns.
* grOogle notes that HCoyote fails the memory test.
<HCoyote> dunno then ... haven't had my hands on 2GB parts for a few months. I try to keep my hands out of the
+ computers now.
<ket> HCoyote - heh.
<grOogle> HCoyote: If they're my chips, they do chipkill.
* ket is still trying to figure out what hte fuck grOogle got his hands on there, because wintec doesn't list it in
+ their reseller listings either
<ket> grOogle - no they don't. i can make out the pns on the supports and they don't support chipkill
<HCoyote> any part numbers?
<grOogle> HCoyote: Hold on.
<ket> HCoyote - the micron numbers are fudged, but hang
<ket> ics A16859AK
<ket> also a DN431090C
<grOogle> http://www.lemis.com/grog/Thumbnails-20050404.html
<ket> HCoyote - hang on i'll link you
<ket> the pair is non-pll fanout yes? i can never fucking remember.. :(
<ket> nonono.. dumbass me. hang on.
<grOogle> ket: BTW, these chips work fine in the server for which they were purchased.
<ket> grOogle - it's probably overvolting the dimms
<HCoyote> what's the problem then?
<ket> grOogle - i have an abit kd7a that does it unintentionally, lets me run midgrade (C2) kingston parts at
+ incredibly tight timings
<ket> HCoyote - well for one that dimm seems to not exist. :P
<grOogle> ket: I checked wiht memtest86+. No info about the voltage or anything else.
<ket> HCoyote - okay, the 13:26 register is the 16859AK
<ket> grOogle - memtest86 is far from foolproof :(
<ket> 43109 is.. making me suffer through a very slow website :P
<grOogle> ket: You said that it would supply info about the voltage and other stuff. It doesn't.
<ket> grOogle - it does SOMETIMES
<ket> grOogle - it does on my netfinity, bp6, and kd7a. it SOMETIMES does on the kt7-raid. if the bios interferes with
+ spd, you're pretty much SOL
<ket> HCoyote - oopsie, i botched the pn. it's 95V857AG, 2.5 wide range freq clock driver; like i said - no chipkill
<ket> WOAH
<ket> http://www.icst.com/icscs/PartSummary.aspx?id=97e3ff13-ab42-459e-bf7c-d736439be249&name=ICS95V857AH&mode=full
<ket> note voltage :P
<ket> also number of outputs.
<HCoyote> so .. again .. what's your issue with the memory?
<HCoyote> it's a 2GB registered ECC dimm
<ket> HCoyote - well now i'm pretty much positive that it's not what it claims. the MT P/Ns don't match private label
+ (which wintec never uses) and the parts just don't match
<ket> HCoyote - dude, micron doesn't make anything below a 256MB in that package for pc3200
<ket> i checked.
<ket> the 13:26 also doesn't go below 210MHz, which i can't imagine is conducive to 200MHz operation
<grOogle> below?
* HCoyote scratches head.
* HCoyote goes to find that ram testing consortium website thing.
<ket> grOogle - right; it's operating range is specifically 210-270MHz
<ket> HCoyote - cmtlabs.com?
<ket> HCoyote - i'm heading there again now, see if they've got fresh
<ket> wintec's been testing w/intel and hasn't certified or even tested a 2GB part, only a 1GB which is listed on AMPO
+ and reseller
<ket> ah, there it is.. 35964741-L, but that's a 2GB PC2700 not PC3200
<grOogle> ket: When did they come up with a minimum operating frequency?
<grOogle> ket: Well, in that range, anyway.
<ket> grOogle - that's the minimum operating frequency of a specific _part_
<grOogle> ket: When did they come up with a minimum operating frequency?
<ket> grOogle - see the square ICS chip in the middle? _that_ part. and that's just how the part works.
* grOogle gives up trying to get an answer to his question.
<ket> grOogle - i've just told you all i know about that particular IC. it runs at a range of 210MHz to 270Mhz with a
+ voltage range of 2.3-2.7V and a nominal of 2.5V
<ket> interesting.. http://www.cmtlabs.com/images/Scans/WTC037.jpg
<ket> that's a pc2700 2gb from wintec.. which has the exact same parts as this supposed pc32700
<ket> er 3200
<ket> ... including the pcb ID
<grOogle> Hmm. That looks familiar.
<ket> grOogle - because it's the exact same dimm.
<ket> it's got different memory chips, but it's the _exact_ _same_ _dimm_
<grOogle> You mean the layout?
<ket> no, i mean it's the exact same part, exactly.
<ket> the ONLY difference is micron chips instead of infineon. other than that, it is identical down to the support
<grOogle> So they can handle the frequencies of both.
<HCoyote> same dimm doesn't mean anything, really.
<HCoyote> We had a few sets of ATP memory at work ... same size ... same speed ... same part number ... one was low
+ profile, the other wasn't.
<ket> grOogle - not necessarily
<ket> HCoyote - btdt (and atp is good stuff of course), but this is different.. there's been a lot of forging going on
+ past six months, encouraged by geils prior forgery
<ket> HCoyote - a LOT of people selling nonexistant parts, overclocking by reprogramming the eeprom or just
+ mislabelling
<HCoyote> um. what wintec part at cmtlabs were you looking at?
<ket> HCoyote - the only 2GB module they've tested, 35964741-L
<ket> http://www.cmtlabs.com/memSearchResults.asp?sManuf=Wintec+Industries&sPN=35964741-L&oSubmit=Search
<HCoyote> that's not the same part as what's pictured in his photo.
<ket> HCoyote - no, it's not by the partnumbers.
<ket> HCoyote - but stickers are easy to forge; look at the electrical component pns. they're absolutely identical.
+ same for support layout.
<ket> HCoyote - pcb numbers _also_ match digit for digit
<ket> Register A16859AK PLL 95V857AG
* ket watches HCoyote compare and go 'oh.. it _is_.'
<ket> grOogle - is accutek your vendor?
<grOogle> ket: I borrowed this RAM.
<ket> sorry, should've said 'a local vendor', and yes they are.
<ket> with a really shitty website
<ket> www.accutek.com.au - so that answers that question
<HCoyote> grOogle: if you have any doubts, just talk to wintec. It's not unheard of for RAM vendors to parts out
+ that aren't listed on their website yet.
<ket> HCoyote - agreed, but i didn't even see it on resale list.
<ket> THAT'S what bugs me
<HCoyote> ket: Tyan has that part listed as a valid part for several of their Intel motherboards.
<ket> it's possible the resale list was out of date, but with all the forgeries and grey market bullshit..
<ket> HCoyote - which part? the pc2700?
<HCoyote> no
<HCoyote> http://www.tyan.com/support/html/memory_s2885.html
<HCoyote> go down to the DDR400 section.
<ket> uhm, s2885 ! intel
<HCoyote> err, sorry
* ket smacks
<HCoyote> http://www.tyan.com/support/html/memory_s5360.html
<ket> well i'll be damned
* HCoyote kicks ket
<HCoyote> STFU.
<ket> ahahah pdp patriot. hahahaha
<HCoyote> so, anyway ...
<HCoyote> grOogle: if you have any doubts about the authenticity, just ask Wintec. I'm sure they'd be able to point
+ you in the right direction.
<ket> hahaha.. ocz.. ocz's micron 512mb passes but their 1gb failed, almost assuredly. pdp is funny though, since
+ they're not pure samsung.. gamble
<ket> HCoyote - images are high enough to do quickie compare on the sticker, it doesn't look forged
<ket> msi just can't push the load, no surprise
<HCoyote> so why did we go on this wild goose chase?
<ket> HCoyote - because i fucking hate forgeries.
<ket> ran into a shithole stickerfactory that was labelling flaky pc66 as pc100 once
<HCoyote> it's likely an OEM'd part.
<ket> considering wintec doesn't do any retail parts? :P
<ket> it's a mass-production resale part
<HCoyote> funny, looks like you can buy it online through wintec's website.
<HCoyote> (wintec ram, in general)
<ket> HCoyote - yah, in antistatic bags or if you buy enough, in a packbox.
<ket> HCoyote - e.g. not retail box
<ket> you don't get instructions, fancy box, how tos, etc
<HCoyote> wow ... 1ecc a second is bad.
<grOogle> HCoyote: Right.
<grOogle> HCoyote: Do you agree with my conclusions?
<ket> grOogle - that the board sucks? of course :D
<HCoyote> That's excessive.
<HCoyote> way excessive.
<ket> HCoyote - aka 'average for msi'
<ket> HCoyote - remember, the board doesn't even have the memory traces correct; 4 dimms to 1 cpu
<HCoyote> the author of bluesmoke was talking to us about testing that ECC was working and suggested we aim a blow
+ dryer at the RAM while stressing the system with memtest ... sounds ilke you may not have adequate cooling
+ around the memory.
<ket> HCoyote - it's not that; it was voltage.
<HCoyote> (and yes, we tested that ... it works)
<ket> HCoyote - known issue with the k8t-master2-far. the fucking thing hasn't got anywhere _NEAR_ enough voltage reg.
<HCoyote> ah
<HCoyote> yeah ... we don't touch MSI.
<grOogle> HCoyote: Could well be. The DIMMs get too hot to touch, a first in my experience.
<ket> HCoyote - yeah, i'm pretty much the most knowledgable about msi here. been working with them since '94
<HCoyote> grOogle: stick a fan over the DIMMS while you're testing and see if that helps.
<ket> HCoyote - to say they've gone downhill is the understatement of the millenium.
<grOogle> HCoyote: I suppose I should, but I've wasted too much time on this already.
<HCoyote> We've had to do that with some lab systems that have 56GB of ram and not adequate cooling.
<grOogle> HCoyote: The DIMMs were only loaners, and the memory I have in there now works fine.
<HCoyote> (we ended up sticking some delta black labels right on the dimms and noticed a drammatic improvement)
<HCoyote> <nod>
<HCoyote> We're only using DDR333 for our opterons. <shrug>
<HCoyote> (except for the few Suns we have)
* HCoyote wants to get one of the HP 2p opteron desktops for home.
<ket> HCoyote - you should be using 2.5V PC3200
<ket> HCoyote - every opteron i've built, i've specced 2.5V. it's dramatically cooler.
<HCoyote> 1) it's a waste of money on older processors 2) it's a waste of money if you have issues with boards that
+ don't support it correctly.
<HCoyote> We don't build systems. We buy them.
<HCoyote> previously, we built 1 and sent it to the local vendor to duplicate.
<HCoyote> now, we just buy tier 1
<HCoyote> On that note, I'm going to go watch people beat each other with sticks.
<quiet1> ... televised british parliment?
