uhh... strange memory problem... more = slower?

[blakerwry]

Hey, I bought some 256MB sticks of Crucial PC2100 to stick into my main PC and my SB51G server(mandrake 9.1)...


Unfortunately 1 of the sticks was DOA.. But the other stick tested good with memtest86 on both these systems.


So, after debating whether to stick the 1 good stick in the main PC or the server, I decided on the server as it would be less hassle and mean less downtime for the server...

So I quit memtest86 and go to boot... and I wait 1/2 hour and it still hasn't finished booting... I figure I must have mucked up something in the BIOS when i double checked RAM timings... uh.. nope... but I reset everything just to be on the safe side and tried booting again... nope... still major delays.

I'm thinking this RAM just passed 4 hours of testing, how could it be causing any problems... so I start downloading new ISOs and prepare to do a reinstall of the OS (thinking a package or 2 I added the other day is causing the problems) before doing a reinstall I feel as though I'm going to kick myself if it's the RAM, so I decide to take out the RAM and try one last boot... low and behold I'm back to ~1 minute boot times. Put the RAM in... 1/2 hour.... take the RAM out 1 minute...


I tried different combinations of RAM and found out that when I have 512MB i'm good... when I have 256MB I'm good.. but when I combine the two I end up with 1/2 hour boot times.... (I should have tried 1GB but didnt think of it)


So, I test the RAM in my main PC and it runs fine with the same combination that refused to work in the shuttle...


So here's where I'm stumped. Why would I have problems running 768MB of RAM in my shuttle SB51G? Is this a Northwood Celeron problem? a SB51G problem? or a Mandrake problem? or just a RAM problem(tried using two sticks of Crucial, a 512MB and a 256MB stick, both PC2100 CAS 2.5)?

I know superCaff runs 2 1GB sticks in his SB51G, but he has a northwood p4 3.06gHz. This makes me think it's not a chipset or a motherboard problem.. but maybe a CPU or OS problem... and since the RAM runs fine in my main PC and tested fine with memtest86 I tend to think it's not a RAM problem... But I can't imagine either Mandrake or p4 Celerons having an incompatibility with 768MB of RAM.


I tried googling, but couldn't come up with anything. Do any of you ahve ideas as to what the problem is and how I can alleviate it?

 

[Howell]

or a Mandrake problem?

Try a Knoppix CD just for kicks.

 

[Fushigi]

My only guess is that it doesn't like having mis-matched memory sizes. Can you try it with 3x256 instead of 512+256?

- Fushigi

 

[blakerwry]

Fushigi:
The SB51G is an intel 845GE based board with support of upto 2GB of DDR 333. Unfortunately, it get's there using only 2 DIMM slots.

At the moment I have 2 512MB sticks, and 2 256MB sticks (not including the non-working one I just RMA'ed)

unfortunately, 1 of my 256MB sticks is ECC and probably won't work correctly in the SB51G.. haven't tried.

This leaves me with a 512MB Crucial stick of PC2100, a 512MB Kingston PC2700, and a 256MB Crucial PC2100.


I believe I've tried the 256MB stick on both slots on the mobo, and when used individually my results are fine. Same goes for the 512MB sticks when used individually.

Mix-matching either 512MB stick with the 256MB stick results in the slowdown.

I didn't try using both 512MB sticks simultaneously as this was not going to be a viable final option for me and I didn't consider it when I was in my troubleshooting mode... I should have. It's possible I have a bad mobo that just doesn't work well when both slots are populated.


Howell:
I was thinking the same thing. I have a spare HDD with an installation of win9x that I can test with if need be.

One thing that did seem odd to me was that the memory testing seemed pretty slow (as in it took over 1 hour to complete a full pass)



In any event I should do more testing. I'm really hoping that I don't have a bad board. An RMA to shuttle is likely to cost me weeks of downtime.

 

[CityK]

When you checked your new stick with memtest, and you thought it was taking a long time (hour for 1 pass), did you have both sticks in together? If so, take the 512 out and retry...bet it works fine i.e. back to speedy operations.

My guess is that its a side affect of the "virtual to physical address space" process that the chipset (NB) performs. Your 512MB DIMMs are probably composed of 16 256Mb (32x8 ) chips where as the 256MB module is likely composed of 16 128Mb (16x8 ) chips. This mismatch in array size is probably causing the chipset some headaches when the CPU sends it an address.

CK

 

[blakerwry]

The 256MB stick is 32x64...

The 512MB of crucial is 64x64

The 512MB of Kingston is 32x64


I did most testing in the SB51G with the two crucial sticks(but I also tried the crucial Kingston combo with the same results)... the Kingston/Crucial combo of 768MB works great in my Soyo.

When I get my other RAM stick back from crucial(hopefully in a week) I'll do more extensive testing.

 

[CityK]

The 256MB stick is 32x64...

The 512MB of crucial is 64x64

The 512MB of Kingston is 32x64

Those are the module density figures you have quoted, as opposed to the chip densities that I quoted above. Unfortunately, Module densities don't convey an accurate picture because of two things: 1) manufacturers don't always follow the same labeling convention and 2) it doesn't tell us how many physical banks the DIMM constitues.

Example: I have a 256MB module that is labeled (in its Module notation) as 32Mx64. This is supposed to mean that the CPU sees the module as 32M storage cells that each contain 64 bits (or 8 bytes, or 1 word). 64 bits is the width of one physical bank (as 64 bits fills the mem bus). Given that each DRAM chip is configured to output 8 bits, you would expect that the module consists of 8 chips (each composed of a 32Mb arrary of storage cells) as 32M*8bits/chip*8chips = 256MB module. However, there is actually 16 chips on the PCB, and the math doesn't add up (would end up with a 512MB module). Also, notice that you would end up with an output of 128bits (16 chips x 8 bits per chip). So the first clue here is that the module is actually composed of two (64bit) physical banks. But that doesn't solve the density issue (you still would still have 256MB per bank). To cut to the chase, what is actually going on is that the DRAM chips are actually 16Mx8 (i.e. 128Mbit chip). Doing the math, we see 16M*8bits/chip*8chips = 128MB per bank, * 2 banks = 256MB.
So, we see that my misappropriately labeled 32Mx64 module is actually 2x16Mx64....meaning that the CPU sees the module as two seperate banks of 16M storage cells, with each cell holding 64bits (8bytes, or 1 word).

Hope that helps. Anyways the point of the matter, is that I can't make out anything without knowing how many chips there are on each DIMM. Given that, and the size of the module, you can pretty much figure out everything.

Intuitively, the Kingston label 32Mx64 can only mean one thing: 16 chips of 32Mx8 (i.e 256Mbit chip), thereby yeilding two physical banks on the DIMM. (Any other way and the math doesn't cut it).

The label on the 512MB Crucial lends itself to speculation, cause we don't know how they follow the labeling convention. For example, you could have only eight chips on that module, and each chip is a 64Mx8 (512Mbit chip....however, this is rather unlikely cause they are so rotten expensive!). There are other configurations it could assume too, but more then likely, it is a 16 chip module, meaning that each chip is 32Mx8 (i.e. 256Mbit chip) and that the DIMM is also composed of 2 physical banks.....It looks like Crucial likes to use the 2x32Mx64 method of labelling their modules.

As for the 256MB Crucial, again inconclusive. If their module labeling is consistent (which there is no way of knowing), then this is actually 16 chips of 16Mx8 (i.e. 128Mbit chip), for a two bank module...ie. 2x16Mx64 in Module notation.....Alternatively, if they aren't consistent, this could be a eight chip module, meaning one physical bank, composed of 32Mx8 chips (i.e 256Mbit chip) in chip notation.... or 32Mx64 in Module talk.

Hi Caromba! Never expected the post to be this long! Anyways, I might be off base (evidenced by your Soyo results), but I still think there is a legitimate possibility that the memory controller (northbridge) in your shuttle is having difficulty with the mismatch in chip array configuration (looks like 32Mx8 for the 512MB modules and 16Mx8 for the 256MB stick).

CK

 

[blakerwry]

Would that be a BIOS probelm or a chipset problem? (I almost find it impossible that a modern day intel chipset would have such a problem)

 

[CityK]

It would be a chipset limitation. However:

I almost find it impossible that a modern day intel chipset would have such a problem

I agree, and the more I think about it, the more I think it shouldn't make one iota of a difference. I think I'm off base. Just speculating (I'm doing a lot of it these days it seems) cause I have seen other people report similar problems when there was a mismatch.

After consideration, it sounds more like an strobe/signal timing problem:

Are you running with SPD on (although it doesn't sound like it given your first post)? Different DIMMs with different settings programmed into their SPD EEPROM might foul things up.

Perhaps you are trying to run one of the DIMMs with too aggressive timings for this board.

Perhaps the traces on "your" board are causing a propagation delay.

CK

 

[blakerwry]

It would be a chipset limitation. However:

I almost find it impossible that a modern day intel chipset would have such a problem

I agree, and the more I think about it, the more I think it shouldn't make one iota of a difference. I think I'm off base. Just speculating (I'm doing a lot of it these days it seems) cause I have seen other people report similar problems when there was a mismatch.

After consideration, it sounds more like an strobe/signal timing problem:

Are you running with SPD on (although it doesn't sound like it given your first post)? Different DIMMs with different settings programmed into their SPD EEPROM might foul things up.

Perhaps you are trying to run one of the DIMMs with too aggressive timings for this board.

Perhaps the traces on "your" board are causing a propagation delay.

CK

i tried SPD and manual, although both crucial sticks are Micron modules on micron PCBs rated PC2100 cas 2.5.

I'm thinking it almost must be something with my particular board... When I get that stick back from crucial I'll have to do some more testing...

 

[Howell]

Those are the module density figures you have quoted, as opposed to the chip densities that I quoted above.

CK

An informative post CityK. Do you know if there is any way to ID a stick when there is no labeling on the PCB and the only infomation I can get is off the chip combined with the number of chips?

 

[blakerwry]

the ways i've gotten RAM info before are:

1) look at the label
2) look at the modules' markings (there's usually atleast a logo, if not product codes)
3) look at the PCB for manufacturer/markings
4) ID the stick using Sandra (don't know if they still include this tool... it was a DOS based tool that came w/ Sandra circa 2000)

 

[blakerwry]

CityK, all my DIMMs are double sided with 4 modules on each side.

 

[CityK]

Do you know if there is any way to ID a stick when there is no labeling on the PCB and the only infomation I can get is off the chip combined with the number of chips?

Sorry Howell, I'm not quite clear what your seeking. Do you mean as in you have a generic module and you wish to ID the manufacturer? If so, you could try a program, like Blake suggested, that will read the info programed into the SPD/EEPROM. CPU-Z has such capability. Caveat - not all manufacturers input such information into the SPD. Alternatively, google with any info (model numbers) you can find on the module. Usually a little white sticker is used (stuck on one of the DRAM chips or on the PCB), and the model number that the manufacturer assigns to the module is written on it. Googleing with that and the chip model number (silk screened on the DRAM chips) will probably get you a hit. Googleing with just the chip model will likely just find you a whole host of other modules also composed of the same chip...

all my DIMMs are double sided

Although you often hear people speak of a single sided or double sided DIMMs, in actuality, there is no such thing as "sidedness" for a module. What is of interest are the number of "physical banks" on the DIMM. Recall that the mem bus is 64 bits, thus to fill it, DRAM chips are organized into 64bit physical banks. Since most DRAM chips usually output 8bits (denoted by *x8 ), you need 8 of them to fill the physical bank. Usually manufacturers will place these 8 chips all on one side of the module. However, there is nothing preventing them from putting 4 chips on either side of the module (or 6 and 2, or 1 and 5...or whateever for that matter - although practicality and design cost would tend to prevent this from ever occuring...except for deep within the bellows of Crucial's most secret labs of "DRAM research gone mad").

So where is this all going? Well, quite simplely, to get a larger (MB) capacity memory module, you can either use DRAM chips of greater density (i.e 256Mbit chips vs. 128Mbit chips) or you could add a second physical bank of the lesser density DRAM chips i.e. using 16 128Mbit chips arranged in two physical banks on the DIMM will give you the same result as using 8 256Mbit chips arranged into one physical bank on the DIMM; that being, of course, a memory module that has a 256MB capacity.

[As a side note, just as a single bank doesn't have to be arranged on one side of the module, the chips of the respective two banks aren't necessarily arranged on opposite sides (ie. all bank 0 on one side, all bank 1 on the other side of the PCB). Instead, you might have 4 of each bank on each side of the DIMM. This is just done to organize the traces and also to minimize their lengths on the PCB, which makes sense given there are 184 pins on DDR SDRAM and 164 on vannila DRAM (ie. could be quite messy, costly, and bad for signal propagation otherwise)].

Today, 256Mbit are commodity DRAM chips. But back a couple of years ago, the decision to go with 16 128Mbit commodity DRAM chips over the "new" and costly 256Mbit chips was a no brainer. The same thing applies to the current situation with costly 512Mbit chips (arranged as 64Mx8 in chip notation).

Now, to address my bolded "usually statement: some DRAM chips output 16 bits per access (denoted *x16). This means, to fill the mem bus, you only need 4 such chips per physical bank. These chips have been used sparingly (more costly cause not as many produced) in most cases, but we're likely to see more and more use of such chips in memory manufacturing.

with 4 modules on each side

You mean chips of course....which would mean that they are of the *x16 variety...However, I find that rather surprising/unlikely. Did you mean 4 chips on the lefthand and righthand of the "whole side" of the PCB (which I could understand some confusion could originate, seeing most DIMMs are arranged with a gap betwen the middle 4th and 5th chips)? or 4 chips on one whole side of the PCB, flip it over and only 4 more chips on the other side of the module?

Sorry, I don't mean to sound like I'm patronizing or anything, but I have no idea what extent you understand the layout of a DIMM. Most people don't. In general, the concepts surrounding the combination of DRAM chips and memory modules are some of the tougher things to grasp in terms of computers - I think this is largely because there are a number of notations and words used that are similar or identical but have a completly different context (i.e. bank - physical kind and internal kind, and the banks on the motherboard) etc., etc. About two years ago, when the mystique shrouding RAM started to interest me, I recall seeing (on the former PC133memory.com forum, that Enhanced Memory Systems had prior to full integration into the Ramtron Mushkin fold) MS from Lost Circuits (he used to moderate it, being an EMS employee) said to one person that was having difficulty grasping something to "not to sweat it, cause it took me almost two years to get the whole picture". Speaking for myself, his assessment is pretty damn accurate. There is a dearth of info on memory, and peicing it together is fraught with confusion. There are, however, several good articles on Ars, Aces, RealWorld and LostCircuits - albiet, none completely comprehensive, but collectively, and after a couple of reads, the pieces do fit together. Be forewarned, cas once (when addressing a RAM question on SR) gave a quote that stated something like "there are few things in life more terrifying then the timing diagram of a SDRAM"....which isn't exactly true, but it does highlight the complexity that surrounds DRAM. I guess that is why it made me smile and has stuck ever since.

CK

 

[blakerwry]

I'm sorry, I must be off today The modules each have 8 chips per side(as in front and back). For a total of 16 chips per DIMM.



To clarify:

I will have 4 DIMMs

1) 512MB Crucial

2) 512MB Kingston

3) 2 x 256MB crucial


On the 256MB DIMMs, the markings on the Micron chips are as follows: 46V16M8 -75 A

so I will assume this means that each chip has an 8 bit interface and has a timing of 7.5ns.

Since these modules are 256Mbytes and I have 16 chips, I guess they are 128Mbit chips... and that the DIMMs are double sided (as in take up two banks)


The crucial 512MB stick also has 16 chips and markings of 46V32M8 -75B. I'll assume that it is also double sided, yet uses 256Mbit chips and should have the same timings as the 256MB DIMM.

I can't see the Kingston very well since it is inside my running case... but I will assume since it has the same 16 chips that it uses 256Mbit chips as well.

 

[CityK]

On the 256MB DIMMs, the markings on the Micron chips are as follows: 46V16M8 -75 A so I will assume this means that each chip has an 8 bit interface

Given what you said above (16 DRAM chips on each of the modules), then yes, exactly.....which is what should be expected. But even further proof is contained in the chip model number itself: 46V16M8....16M8 is just another notation conveying the same thing as 16Mx8.

[In otherwords, its a 128Mbit chip - it has 128M one bit storage cells that are arranged, internally, into 8 arrays (*x8 ) of 16M (16Mx*), where each arrary outputs 1 bit of information per access, for a total of 8 bits per chip per access......So, pretend for a second, that this one DRAM chip constitutes a memory module. In this case, the CPU sees 16M storage cells, which it believes contain 1 byte each....

...Building upon this concept, lets take a DIMM composed of 8 of these DRAM chips (i.e. it fills a single physical bank). The CPU still sees a row of 16M cells, but this time it believes each cell contains 8 bytes (the combined output of the all the DRAM chips)...

..Again, taking the concept further to the actual case we have here with your dual banked DIMM: The CPU sees a 32M long row (2 x 16M) of storage cells that it believes hold 8bytes each...

...So when the CPU is accessing (reading/writing) something in the RAM, it hands the memory controller (northbridge) a "virtual" address which the chipset then has to translate into the actual physical address.....The same thing applies to HDDs - the CPU has no knowledge of the actual physical geometry of the disk or where a byte of information is actually stored....or to draw from real life, I'm pretty sure Tannin doesn't give a rats ass where in the warehouse his supplier stores Gigabyte boards (they could be underneath a pile of ECS - all too Tea's shock and horror), just as long as when he places a order, he recieves his request when he expects it to arrive back in the shop.]

and has a timing of 7.5ns

Yep, clock cycle time, tCK, 7.5ns (1/133MHz).

Since these modules are 256Mbytes and I have 16 chips, I guess they are 128Mbit chips... and that the DIMMs are double sided (as in take up two banks)

Yep...er, just make the last part: and the DIMMs are dual banked (and happen to have DRAM chips on the two sides of the PCB) :wink: .... You could also say that the DIMM takes up two banks too; each DIMM socket on the motherboard has traces running to it to address 2 physical banks of memory if need be. If you look at a program like H.Oda's WPCRedit, you'll see that when you adjust the memory timing settings, its done in terms of banks 0&1, 2&3, ...

The crucial 512MB stick also has 16 chips and markings of 46V32M8 -75B. I'll assume that it is also double sided, yet uses 256Mbit chips and should have the same timings as the 256MB DIMM

Sounds right, with only the same small subtle, but important, correction as pointed out above - the DIMM is a dual bank module; the fact that it has DRAM chips on both sides stems only out of consequence of there being a finite amout of real estate on the single side of a PCB - ie. the DRAM chips' physical size, and the number of traces running to them on the PCB, percludes them from all being placed on one side of the PCB.

Kingston... but I will assume since it has the same 16 chips that it uses 256Mbit chips as well.

Sounds right to me.

CK

 

[CityK]

Again, taking the concept further to the actual case we have here with your dual banked DIMM: The CPU sees a 32M long row (2 x 16M) of storage cells that it believes hold 8bytes each...

Incidently, you can see from this how module density figures tie in with all of this: 32Mx64 (also denoted as 32M64)....which of course mathematically gives you a 256MB capacity

However, from the context of looking at the DRAM chips themselves, you can see that we would have know way of knowing that these were 128Mbit chips from the module density figure, without also knowing that there are 16 DRAM chips on the module.

CK

 

[blakerwry]

I saw on Tyan's site that they have certified the 256MB DIMMs as being completely compatible with one of their 845E boards.

I wonder if it's possible the onboard gfx are interfearing (845GE uses "shared" memory) with normal operation... I don't know how the 845GE's graphics work, but I'd assume that they'd be assigned a contiguous block of RAM from either the 1st part of the 1st DIMM or the last part of the last DIMM. Either of which I would assume wouldn't cause problems when using two different types of DIMMs.

 

[Howell]

Do you know if there is any way to ID a stick when there is no labeling on the PCB and the only infomation I can get is off the chip combined with the number of chips?

Sorry Howell, I'm not quite clear what your seeking. Do you mean as in you have a generic module and you wish to ID the manufacturer? If so, you could try a program, like Blake suggested, that will read the info programed into the SPD/EEPROM. CPU-Z has such capability. Caveat - not all manufacturers input such information into the SPD. Alternatively, google with any info (model numbers) you can find on the module. Usually a little white sticker is used (stuck on one of the DRAM chips or on the PCB), and the model number that the manufacturer assigns to the module is written on it. Googleing with that and the chip model number (silk screened on the DRAM chips) will probably get you a hit. Googleing with just the chip model will likely just find you a whole host of other modules also composed of the same chip...
CK

Actually It's worse than that. I have no motherboard that will give me the same amount of RAM. There is no silkscreen or sticker on the PCB to help with identification. At some point I wrote that they were PC100 but I can't get them to work with modern boards. Googling the chip numbers returns no hits.

 

[Howell]

I have no motherboard that will give me the same amount of RAM.

What I meant to say was that different motherboards ID the same stick as different sizes.

 

[blakerwry]

Just found this BIOS update from Shuttle:

FB51S034
Date code: 05/09/2003 Checksum: 3600

1. Adjust fan setting.
2. Fixed install two different size of DDR , system is not stable.3. Added HT CPU to show "HT Enable" when BIOS post.
4. Added new display string for DVD+CD-RW devices.
5. Support P4 D1 Core Stepping CPU. (Northwood)

 

[CityK]

So did you flash the new BIOS and retry the mem?

CK

 

[blakerwry]

na... I haven't had a chance to burn the BIOS to CD.

Right now I'm using 2x256MB sticks of Crucial so it doesn't effect me. When I reboot next I'll flash to the new BIOS though.