Memory

On some systems, you can interrogate the memory and find out how it's organized (so, you plug it in, then run the system-info and it tells you how the memory controller has configured the RAM).

On MacOS, "System Information.app" gives Size: 4 GB Type: DDR3 Speed: 1067 MHz Status: OK Manufacturer: 0x857F Part Number: 0x483634353155363446373036364700000000 Serial Number: -

on each module, it's fairly easy to see if the part numbers and manufacturers match. On Windows, I've used DOS-prompt utility wmic

wmic memorychip get manufacturer,serialnumber

with variations on the what-to-get string...

So, your box of memory can be plugged into the machine and (after rebooting) it'll be possible to explore your options. The "organization" gets complicated, on memory chips; some firmware will recognize a set of banks-and-sizes combinations, then a newer generation of chips, in the same module form, will add combinations that are unsupported. It's even possible for a firmware update to cause a perfectly good RAM module to be rejected.

If you're bold, do the experiment. If not, most memory merchants market matched sets for precisely this situation.

I'm not concerned as to how *I* get to a working system (my kernels tell me what the spd's indicate as part of the dmesg output).

My question goes to addressing the flaw? in the documentation that would confound an "IT guy". I.e., how does he know that a "set of four" (or, a "set of two"!) are "sufficiently matched" to AVOID the "slight reduction in performance" that "by two" matching would incur -- given that they only present "size, rank and organization" as the criteria (in both cases).

Do these folks buy "new sets" every time they want to upgrade memory in systems? Do they take the (presumably matched?) old sets and wrap a rubber band around them to ensure they travel together (as a matched set) for FUTURE uses? Or, do they just discard them??

The "IT guy" generally won't take chances, he'll get a memory merchant to make the recommendation, and usually buy a boxed set. That's sometimes wise, but the overhead is a problem for us stingy guys.

I've seen computer manufacturer data that was wrong, and memory merchants that knew better. I've seen firmware updates reject useful memory modules for lack of a (non-specified) feature. I've heard of firmware updates being necessary for some memory installs. I've seen memory specified as ONLY to be replaced in matched sets (i.e. if you have a known-bad part, you have to pull all of the set, and replace all of the set).

Industry part number schemes are, possibly intentionally, confusing as to what exactly the inner character of the module is. Motherboard manufacturer data usually IS correct, but sometimes grows stale as memory manufacture progresses.

You can always mount them and then see what SPD finds. What it says is generally true that putting odd sized ones in will be the worst possible performance (might not even work on some chipsets) pairs are better and on some boards quads are (slightly) better still.

If you care about ultimate memory bandwidth performance then you try hard to meet these criterion, but it can be expensive to get that last

Superfast (aka gaming) ram is generally not worth the extra money.

What tends to happen if they are the same size, rank and organisation but different speeds is that the system runs to match the slowest (rather than not running at all).

I prefer pairs or quads from the same maker and batch that way you don't get any nasty surprises.

SPD will tell you enough but otherwise write a memory intensive program or run a well known memory speed benchmark program to verify results.

He doesn't and he doesn't care. So long as the thing boots after a memory upgrade a modern PC will be more than fast enough to do general office functions even on hopelessly mismatched sets of ram. And I have never known adding good ram to slow a machine down it just might not speed it up as much as an optimal memory configuration would.

The IT guys I know who did this sort of thing hang onto anything good still young enough to be useful and use good old rubber bands and static protect bags to keep sets together (rather than a junk drawer).

They are also a lot smarter than you are giving them credit for. My brother in law used to do this sort of work for a major university until he retired. He is a recognised authority on Hackintosh as a hobby.

A lot of big corporate places don't upgrade PCs that much, they buy new ones from scratch. Universities and smaller businesses might, but in most cases they won't even notice a 5% drop in ultimate performance compared with the speed up that comes from having more working memory. It might be a little bit faster still but who cares, most desktop PCs are already way over powered for the work they have to do.

You're side-stepping the issue.

If what you say is true, then why even BOTHER to print these constraints? Why not just advise the user to pick memory that has less stringent constraints that the hardware and BIOS are willing to ACCOMMODATE and gloss over the performance references? If the user can't TELL that this set of two sticks is slightly different from the other pair IN THE SET OF 4 based on observations of performance, then why bother him with this detail?

I.e., if the detail is significant, then present it in a clear and definitive manner (instead of mealy-mouthed words)

So, *my* approach coincides with theirs...

I'd be curious as to where you see my claim that they AREN'T smart?

Again, you're missing my point. Explain why the manufacturer published the descriptions of the "replacing memory" constraints that I quoted. Clearly, they feel that *someone* cares about these details. My point is that the details they provide are ambiguous wrt "pairs" vs. "sets-of-four".

If they intend these to be "buy a PREPACKAGED set of 2 (or 4) matched DIMMs" (i.e., let the memory manufacturers sort out what needs to be "matched" in the devices), then the comments about rank, size, etc. are superfluous.

Because none of the customers IS clear on memory characteristics; FPM versus EDO versus SDRAM, PC100, PC133, ECC, buffered, 'fully buffered', .... the list of not-quite-orthogonal labels for memory strategies is very large. Size and shape, too. And power supply voltage. And latency differences, clock speeds...

The server manufacturer wants you to buy the RAM from HIM, and the warranty requires it. After warranty is up, you can continue to be a customer or... or the manufacturer is just fine with you being lost in the cold.