single sided vs. double sided, how to tell?

How do I know, by vendor description, if a RAM stick is single sided (chips
only on one side) vs. double sided?

Are there advantages to getting single sided vs. double sided?

TIA

~signmeuptoo

 

In most cases, there is too little info available, to know what
you are buying. Especially with the cheaper generic RAM.

With Crucial, you take the Crucial part number, find a picture of
the Crucial module on Newegg, get the Micron part number off the
picture of the sticker, and then download a datasheet from Micron
(they have a section devoted to their modules and to the chips
they make). I have never seen a cross reference table, that
converts a Crucial part number, into the Micron module used
for that purpose.

With Kingston, datasheet download links are available on the
product description page.

For Corsair, you can go to corsairmicro.com and download a
datasheet from their product listings.

Samsung also has datasheets:
http://www.samsung.com/Products/Semiconductor/common/product_list.aspx?family_cd=DRM030202

Those are all the ones I know of, or have searched for.

Crucial gives a hint, by putting "-8T" or "-16T" on the end
of the part number, but not all adverts include that info.

In terms of construction, a bank of RAM is any number of
chips put side-by-side, to build a 64 bit wide array of memory.
The most popular widths of memory chips are 8 bits and 16 bits.
That means you could construct an array with four or eight
chips. (It is irrelevant, but I believe I've seen 2,4,8,16, and
32 bit wide chips, but some of those are used for more obscure
purposes. I have a x32 chip, for example, on a Xilinx development
board.) When you find 8 chips on a module, it could be (8) (x8)
chips or it could be two groups of (4) (x16) chips.

So, what is relevant about banks of RAM ? In the case of FPM, EDO,
SDRAM, DDR, DDR2, all the chips on the module load down the address
bus. A module with 16 chips puts 16 loads on the address bus.
A module with 8 chips puts 8 loads on the address bus. Each load
has a certain input capacitance. A capacitor stores
electrical charge, and the more you've got, the harder it is to
wiggle the signal on the line up and down. What that means, is
a DIMM with fewer chips on it, can be run at a higher speed if you
are overclocking. So, what you really want, is a module with as
few chips as possible, for the highest operating speed. But of
course, a module with few chips on it, gives you less total memory.
For overclockers, two single sided 256MB DIMMs, run in dual
channel, is a typical overclocking config, but few people
are really happy with only 512MB total RAM. If RAM chips of
32Mx16 were used, you could build that module with four chips.

Registered memory gets around this, as a register (buffer) chip
on the DIMM, forms an intermediate stopping point for the signals.
The address bus from the motherboard only drives the register chip,
then the register chip drives the 8 or 16 or 32 memory chips. The
register chip increases the latency in getting to the module, so
peformance drops due to that, but the advantage is, that many more
DIMMs can be driven, before the address bus runs out of drive
strength. That is why server boards can have larger arrays of
DIMMs. But, as an overclocker, I doubt a registered DIMM would
make a good candidate for overclocking, as the register chip
has frequency limits to it that are likely tighter than
a memory chip.

At the densities most people are interested in, modules will
have eight or sixteen chips. Now that larger chips are available,
it is possible to find single bank eight chip 512MB modules,
which cause less loading than a double bank sixteen chip 512MB
module (that is the commodity configuration). Usually the
price will be an indicator of what you are getting.

For example, the first Crucial module here, uses 64Mx8 chips,
and the second module uses 32Mx8 chips. Per 512MB worth of
memory, the first module is 50% more expensive. If the first
(1GB) module was half populated, with only eight chips, it
would give 512MB of RAM, with half the electrical load on
the address bus, of the second module.

1GB ‹ CT12864Z40B DDR PC3200 CL=3 NON-ECC UNBUFFERED $181.99
512MB ‹ CT6464Z40B DDR PC3200 CL=3 NON-ECC UNBUFFERED $ 59.99

In order to buy and use 512MB single sided DIMMs, your motherboard
manual has to claim support for 1GB modules. This has to do with
support for the memory chip size, rather than the module itself.
As the chips are the same 64Mx8 on a 512MB single sided and a 1GB
double sided DIMM, the row and column address signals are the
same.

I hope that isn't too confusing. With all the rows, columns,
banks, and ranks on DIMMs, it is hard to keep track of all
the issues.

Paul

 

<SNIP>



Jokes aside, that was an excellent description.

Ben

 

Paul, could you please explain to me the ##Mx## thing? I mean, for
instance, is a 64Mx8 mean there there is 8 chips of 64 Meg capacity? And
then, does that description skip showing the X1 or X2 banks. Why then
isn't memory described as, for instance 64Mx8x1 to designate a 512Meg stick
and 64Mx8x2 used to describe a 1Gig stick? Why do they omit that last
number?

Am I correct in understanding what you are saying?

Also, I know that DDR memory runs on both the rise *and* fall time of the
clock cycles, but how does DDR3 of the latest Video RAM work?

 

64Mx8 means the silicon die has 512 megabits of RAM. A parallel
bus of 8 bits in width, is used to read and write the memory in
the chip. So, logically speaking, think of the memory as an
array 64 million locations deep and 8 bits wide.

Inside a typical memory chip, memory is broken into four arrays.
Each array is "rows" x "columns" wide. The four separate
arrays of memory are "banks". So, from a terminology perspective,
that leaves "ranks", to describe the two banks of memory on a
double sided DIMM.

The internal organization of a memory chip, is only important
when trying to figure out how big a memory chip that a
Northbridge can support. Northbridges (or the integrated memory
controller in an Athlon64/Opteron) have a limited number of
address bits, to drive multiplexed row and column addresses to
the memory chips. And, that is where the "density issue" comes
in - when you can only see half the memory on a DIMM, it
can be caused by a shortage of address bit from the memory
controller interface.

As for describing memory, a description of (16) 64Mx8 would be
enough info to conclude a DIMM was double sided (two ranks). For
reasons unknown to me, vendors refuse to describe it that way.
After all, it is in their best interests, to provide as little
info as possible about the shortcomings of any product.

This will answer all your questions about GDDR3. One of the
advantages of GDDR3, is reduced power consumption, due to
the modified I/O used, and that is one of the reasons to look
for it on a video card:

http://download.micron.com/pdf/pubs/designline/dl402.pdf

Paul

 

I think I understand, maybe. I took digital electronics in tech school,
but didn't study RAM, didn't go that far, due to what they were training
us.

So what you mean is that there are 64 million transistors and the buss that
transferrs the data to and from them is 8 bit across? So why is it that it
so happens that 64 times 8 comes out to 512? Is that just a coincidence,
or are there 8 different sets of 64 million transistors?

Is there a site or video that you like to use to explain all of this?

I am feeling pretty dumb right now, because I thought that I used to know
how this works and I don't remember now.

Also, do you happen to know how many layers deep these transistors go, and
is that part of the equation?

Would this mean that 128Mx8 mean that it was a 1 gig module? And then, if
there was 64Mx4 it would be a 256Meg module? Would this second example use
four chips?

 

Memory Basics (slide show with voice)
http://www.corsairmicro.com/memory_basics/153707/index.html

 

[snip]

No.

Those designations say *nothing* about the physical layout (i.e., number of
chips) of a SIMM or DIMM. For all intents and purposes, they are purely
logical descriptors. In the case you cited, it means either of two things,
depending on the context:

1. - In the context of memory chips themselves, "64Mx8" would indicate a total
of 512Mbits of memory, arranged as eight "columns" (or "rows", if you prefer;
but this is starting to get into terminology more properly used when
discussing the inner workings of the RAM chips themselves -- refresh cycles,
pre-fetch delays, etc. -- where the context is somewhat different), each
64Mbits deep. The "end point" of each of these eight "columns" is brought out
to one of the pins on the chip; so that is how much data can be accessed at
any given moment. Which "row" of data is presented at any given moment is
determined by the memory controller via the address bus (and by convention, if
not absolute necessity, these are kept "in sync" for all of the chips which
form the array). Hence, for a standard 168-pin DIMM, which by definition
provides a 64-bit (or 72-bit, if an ECC/parity type) data bus, you'd need
eight (or nine) of these chips

2. - In the context of SIMMs/DIMMs, all of the memory on that SIMM/DIMM
(regardless of how many chips it might sport) is arranged "in aggregate" as
eight "columns" (again, or "rows") of 64Mbits each. (Side note: This
particular example would be an extremely odd SIMM/DIMM indeed; AFAIK, the
largest "x8" SIMMs topped out at 16MB -- and even those were rare -- and there
were never any "x8" DIMMs. But this is beside the point, really.)

The key thing to understand is that in both cases, the (much) smaller number
represents the number of bits which are *simultaneously* available (on the
data bus) to the memory controller. Which means that the other (much larger)
dimension represents the length of those "columns". And this is where "single
bank" vs. "double bank" (sometimes called "single-sided" and "double-sided",
respectively; but this incorrectly implies that the physical placement of the
chips on the PCB is necessarily significant) gets into the act...

The memory controllers integrated onto the motherboards of PCs can be (and
often are) significantly limited in terms of how "long" a "column" they can
support (there are some arcane technical reasons for this; don't sweat it for
now). That "512Mx8" example you mention is a rather long column, by
not-so-long-ago standards; and a (still hypothetical, I think -- but I've not
checked lately) 1024Mx8 chip would represent a column length supported only by
some of the very latest chipsets. So in order to provide "bigger" DIMMs,
memory manufacturers sometimes double up the arrangement, effectively putting
two DIMMs together into a single physical device. To you, it *looks* like a
single DIMM; but from the computer's (specifically, the memory controller's)
point of view, it's really *two* DIMMs, each half the size of the whole.
Where this *can* get you into trouble is, those memory controllers are (at
least typically) also limited in terms of how many memory *banks* they will
support. So putting in a single so-called "double-sided" DIMM eats up two
"banks" of the controller's capacity. Depending on the mobo design, you can
run out of memory controller "banks" before you run out of physical DIMM
slots.

Clear as mud?

--

Jay T. Blocksom
--------------------------------
Appropriate Technology, Inc.
usenet02[at]appropriate-tech.net

"They that can give up essential liberty to obtain a little temporary
safety deserve neither liberty nor safety."
-- Benjamin Franklin, Historical Review of Pennsylvania, 1759.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Unsolicited advertising sent to this domain is expressly prohibited under
47 USC S227 and State Law. Violators are subject to prosecution.

 

[snip]

Not transistors. Basically, little capacitors.

[snip]

Uhhh... Because arithmetic works that way?

[snip]

It's not a coincidence. It's math.

[snip]

I suspect that part of the problem is that you are confusing yourself via
"information overload" -- possibly exacerbated by some incorrect preconceived
notions.

[snip]

"Layers", per se, aren't really significant. You would do well to stop
worrying about the *physical* construction of the memory chips and such, and
concentrate on their logical attributes. In that sense, the "arrangement" of
any given memory chip can be usefully thought of as a big (and usually very
lopsided) two-dimensional array.

[snip]

First, *IF* that term was being used in the context of a memory "module"
(i.e., a DIMM or SIMM), then it would indeed indicate the capacity of that
"module". But more typically, that term would be used to describe a memory
*chip*.

Second, your math is near-certainly off (at least the "x8" part nearly always
denotes *bits*, not bytes). Hence, that would be a 128MByte chip.

[snip]

Again, this depends (almost purely) on the context; and the provisos regarding
your math still apply. In this case, the most likely interpretation is a
32MByte chip (four "columns" of 64Mbits each).

You mean if we were talking about a SIMM or DIMM? Then... maybe, or maybe
not. But be aware that, AFAIK, there has never been any such thing as a
"64Mx4" SIMM or DIMM (at least none that are applicable to anything resembling
a PC). Bottom Line: The number of physical chips does not necessarily have
*anything* to do with the "logical arrangement" (or capacity, or much of
anything else) of any given SIMM or DIMM.

See my other f'up to you for more of the gory details.

--

Jay T. Blocksom
--------------------------------
Appropriate Technology, Inc.
usenet02[at]appropriate-tech.net

"They that can give up essential liberty to obtain a little temporary
safety deserve neither liberty nor safety."
-- Benjamin Franklin, Historical Review of Pennsylvania, 1759.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Unsolicited advertising sent to this domain is expressly prohibited under
47 USC S227 and State Law. Violators are subject to prosecution.

 

I think the URL that Ed gave, is a good introduction. The only
improvement I can see in this presentation, is if they dropped
the cheesy voice over. I find it takes too long to skim through
the info in the presentation. You'll have to be patient to
listen to all of this:

http://www.corsairmicro.com/memory_basics/153707/index.html

It is easy to get the "eights" in all the math mixed up -
I've even found errors counting bits and bytes on the web
sites of the major memory manufacturers.

The reason we are concerned with the internal arrangement of
the memory chips, is to determine whether a Northbridge
can interface correctly to them or not. The number of
bits needed to address a row and a column inside the memory
chip is important when figuring out whether a new denser
memory chip will work with an old Northbridge.

Page 16 of the Corsair presentation shows a picure. It
shows four "slabs" made from 4096x1024x8 bits. That is
a total of 4 megabytes in a slab. The four slabs or banks
means the total memory for the memory chip is 16 megabytes.
Now, all the memory is sitting flat on the chip, as seen
in this presentation (this picture is for a different
capacity of chip).

http://www.lostcircuits.com/memory/eddr/2.shtml

If you view the chip as a black box, it is a rectangle
16 million locations on one edge and 8 bits wide on the
other edge. I view the memory that way, because the
memory is a "by 8" or "x8" device, and as you saw with
each slab, every operation there involves a certain
memory address and the 8 bits that reside there. If
the chip was a x16, then each read or write to a location
in one of the slabs, would change the 16 bits stored at
that address.

Now that we know the total capacity of the device
(by multiplying rows * columns * bits_per_address * banks
and dividing by eight_bits_per_byte to get bytes), we
can count the memory chips on the DIMM and work out the
capacity of the DIMM. If we had (16) 16mx8 chips, that
is a total of 256 megabytes of memory.

I think the Corsair slides make this easier to understand.
USENET doesn't offer a chance to do good diagrams, and
the Corsair slides give you something to look at.

As for how the memory is constructed internally, I haven't
a clue how they do it any more. That lostcircuits article
gives some ideas.

Paul