Z420: Memory read error - DB parity error

[Robert2222]

Hello,

on the mainboard of my Z420 workstation there are 8 DIMM slots.

Before my memory upgrade, 4 of these slots (the black ones -> 1,3,6, 😎 were occupied by:

Samsung 4GB 2Rx8 PC3-12800E-11-11-E3

Part Number: M391B5273DH0-CK0

I have bought new memory (4 x 4GB) of exaclty the same type:

Samsung 4GB 2Rx8 PC3-12800E-11-11-E3

Part Number: M391B5273DH0-CK0

I have inserted the new memory into the available white DIMM slots (column A in the picture below). After that memtest86+ fails with this message:

929-Fatal MCA error.

    HA error detected CPU 0 DIMM Slot 5 or 6

    Memory read error - DB parity error

I tried many sorts of configurations by re-seating the memory modules into different slots (check columns C to G). And the funny observation is: whenever I put one of the OLD memory modules into a channel where both slots of the channel are occupied this error comes up (columns E, F and G support that observation).

 

How is this observation even possible?

All modules are perfectly fine as shown by the tests in column C and D! And they are exactly the same modules as indicated by the specs!

Regards,

Robert

[Robert2222]

push

[mdklassen]

From: http://h10010.www1.hp.com/wwpc/pscmisc/vac/us/product_pdfs/Z420_datasheet-highres.pdf

8 DIMM slots, Up to 64 GB ECC unbuffered DDR3 1600 MHz; 4 channels per CPU

Each processor supports up to 4 channels of DDR3 memory.

To realize full performance at least 1 DIMM must be inserted into each channel.

This last sentence explains why your columns "C" and "D" work -- the inserted RAM follows that recommendation, namely exactly 1 DIMM per 1 channel, for each of the 4 channels.

Have you tried "new" RAM in slots 1/2/3/4, and "old"   RAM in slots 5/6/7/8 ?

Have you tried "old"   RAM in slots 1/2/3/4, and "new" RAM in slots 5/6/7/8 ?

The motherboard uses "dual-channel" for slots 1 & 2 together, and 3 & 4, and 5 & 6, and 7 & 8.

That's why MEMTEST reports "5 or 6"  -- one pair of the sticks is "bad".

Also, run MEMTEST with "new" RAM in 1/3/5/7, with none of the "old" RAM installed.

Then, move the sticks: 1 -> 5, 3 -> 7, 5 -> 3, and 7 -> 1, and test again.

Obviously, MEMTEST occupies some of the RAM being tested, and cannot test that RAM.

[Robert2222]

Have you tried "new" RAM in slots 1/2/3/4, and "old"   RAM in slots 5/6/7/8 ?

Have you tried "old"   RAM in slots 1/2/3/4, and "new" RAM in slots 5/6/7/8 ?

I'm testing that right now but I'm very sure this will fail because from columns E to G it seems like I get a parity error every time that I seat an old module into a fully loaded channel.

Also, run MEMTEST with "new" RAM in 1/3/5/7, with none of the "old" RAM installed.

Then, move the sticks: 1 -> 5, 3 -> 7, 5 -> 3, and 7 -> 1, and test again.

You're supposed to load the black slots first in every channel according to the Load Order in the manual. So loading 5+7 (white) without loading 6+8 (black) doesn't make much sense to me, but maybe I'm overseeing something - what's the reasoning behind this one?

[mdklassen]

> I'm testing that right now but I'm very sure this will fail

To me, it makes sense to first run the experiments, and then to analyze the results, to try to make sense.

So, many hours later, what was the result from this experiment?

> I get a parity error every time that I seat an old module into a fully loaded channel.

You have 8 slots, but only 4 channels. 

To me, it makes sense to fill a channel, e.g., 1/2, with two "old" sticks, or with two "new" sticks;  definitely not one "old" and one "new" stick.

With 4 sticks, each of the 4 channels runs in "single-channel" mode.

With 8 sticks, each of the 4 channels, e.g., pairing 1 with 2, and pairing 3 with 4, and pairing 5 with 6, and pairing 7 with 8, runs in "dual-channel" mode.

In "dual-channel" mode, when 8 bits is written to a pair of sticks, each stick receives 4 bits.

Both sticks receive their 4 bits at the same time. 

This is faster than just one stick receiving all 8 bits, while the other sticks have nothing to receive.

MEMTEST should be showing that the RAM is running in dual-channel mode.  Correct?

MEMTEST reports that the 5/6 "pair" has a problem, because MEMTEST cannot "override" the dual-channel mode, to diagnose whether the problem is with '5' or with '6'.

>  You're supposed to load the black slots first in every channel according to the Load Order in the manual.

Yes, in that arrangement, each channel is "half-full", and that's why columns 'C' and 'D' work properly, as you have indicated.

> Also, run MEMTEST with "new" RAM in 1/3/5/7, with none of the "old" RAM installed.

> Then, move the sticks: 1 -> 5, 3 -> 7, 5 -> 3, and 7 -> 1, and test again.

> So loading 5+7 (white) without loading 6+8 (black) doesn't make much sense to me,

What colour are the 8 slots?  Is it:

1. Black

2. White

3. Black

4. White

5. Black

6. White

7. Black

8. White

or is it:

1. Black

2. White

3. Black

4. White

5. White

6. Black

7. White

8. Black

???

> but maybe I'm over looking something - what's the reasoning behind this one?

Because column 'C' works, I am thinking that 1/3/5/7 all are "black".  Correct?

Does MEMTEST report the specific memory-address that is "failing"?

The value should help identify the specific channel that responds to that address.

[mdklassen]

Note that column 'G' works, when the 1/2 pair has two sticks of "new" RAM in the first channel, and there is only 1 stick per channel for the other 3 channels.  This implies that the first channel is running in "dual-channel" mode, without any problems, when you do not mix "old" and "new" in the same channel.

[Robert2222]

I colored the rows of my Excel screenshot keeping the colors on the mainboard in mind - sorry I should have mentioned that 🙂 So from the screenshot you can see that slots 1,3,6 and 8 are black.

I have made a lot of new tests, starting with your suggested tests in column H and I. Additionally I have added in a second row in which channel the parity error arises (notice how in columns H and I the parity error is always where the old modules are):

 EDIT: That's a small picture, I didn't find out how to make it bigger but if you zoom in with ctrl + mouse wheel you can see everything.

> This implies that the first channel is running in "dual-channel" mode, without any problems, when you do not mix "old" and "new" in the same channel.

That's almost what I wrote in my first post:

> And the funny observation is: whenever I put one of the OLD memory modules into a channel where both slots of the channel are occupied this error comes up (columns E, F and G support that observation).

So when one of the old modules is part of a channel which is FULLY LOADED (meaning e.g. 1+2 both filled, or 3+4 both filled ecetera ... no matter if it's old+old or old+new), I get the parity error. If you compare that observation to my new screenshot, you will notice that this is true in EVERY column. But I do not understand this observation: why would be old memory modules work alone (all green columns!) but not when paired (all red columns!)?

[mdklassen]

> So when one of the old modules is part of a channel which is FULLY LOADED (meaning e.g. 1+2 both filled, or 3+4 both filled

> ecetera ...   [et cetera ???]

> no matter if it's old+old or old+new), I get the parity error.

Hmm.

In the very first table that you posted, column 'G' had 1+2 fully loaded with "new" RAM, and you reported that it worked.

Now, you are reporting that when 1+2 are fully loaded with "new" RAM, you are getting an error-message.

Since everything worked in that Column 'G', I would remove any "old" RAM from 3+4, and fully load 3+4 with "new" RAM, and test again.

If that test is OK, I would remove any "old" RAM from 5+6, and fully load 5+6 with "new" RAM, and test again.

If that test fails, I would remove the "new" RAM from 5+6, and try another pair of "new" RAM in 5+6.

If that test fails, I would remove the "new" RAM from 5+6, and try a pair of "old" RAM in 5+6.

> If you compare that observation to my new screenshot,

Sigh.  Not yet visible to me -- I must wait for a moderator to "approve" the posting of that image.

> I do not understand this observation: why would be old memory modules work alone (all green columns!) but not when paired (all red columns!)?

My conjecture is that "old" RAM should not be paired with "new" RAM, no matter how "identical" they claim to be.

What happens with "new" RAM in 1+2+3+4, and "old"   RAM in 5+6+7+8 ?

What happens with "old"   RAM in 1+2+3+4, and "new" RAM in 5+6+7+8 ?

Does MEMTEST give the address of the RAM (some value between 0 and 8GB) that is "failing" ?

Does it identify the channel, e.g., '5+6', that is failing?

Is it always '5+6' channel that fails ?

[Robert2222]

> ecetera ...   [et cetera ???]

Oops yes, sorry for the typo, I'm not a native speaker 🙂

> In the very first table that you posted, column 'G' had 1+2 fully loaded with "new" RAM, and you reported that it worked.

Now, you are reporting that when 1+2 are fully loaded with "new" RAM, you are getting an error-message.

Uhm no, column G is still the same? It's working when 1+2 are loaded with "new" RAM? Maybe your eyes skipped to another column in the results-row? Sorry that it's so tiny, you should zoom in with ctrl + mouse wheel 🙂

EDIT: Okay now I understand from which part of my previous message you made that conclusion. I wrote that 1+2 fails "no matter if it's old+old or old+new". I meant that 1+2 ONLY fails at old+old or old+new, it does NOT fail at new+new.

> Sigh.  Not yet visible to me -- I must wait for a moderator to "approve" the posting of that image.

Ohhhh that's why, okay. Then we should wait until you can see it.

> Does it identify the channel, e.g., '5+6', that is failing?

Yes, I added that information. We should really wait until you can see the new screenshots 🙂

> Is it always '5+6' channel that fails ?

No ... you will see it on the screenshot

> Does MEMTEST give the address of the RAM (some value between 0 and 8GB) that is "failing" ?

No, unfortunately not.

[mdklassen]

>>> whenever I put one of the OLD memory modules into a channel where both slots of the channel are occupied

>>> this error comes up (columns E, F and G support that observation).

I still cannot view the "new" table. 

From the "old" table, in column 'G', you reported that it "works".

Now, you are saying that you get an error in column 'G'.

I am confused!

> Uhm no, column G is still the same? It's working when 1+2 are loaded with "new" RAM?

No, you stated this error comes up (column ... G ...

Is that a "mis-statement" on your part?

> Maybe your eyes skipped to another column in the results-row?

No, I did not.

> Sorry that it's so tiny, you should zoom in with ctrl + mouse wheel

I still cannot view the "new" table -- still waiting for a moderator to "approve" the viewing of the picture.

We should really wait until I can see the new screenshot.