🚀 go-pugleaf

In article <1994Apr11.062202.14307@selway.umt.edu>,
which was posted to comp.sys.ibm.pc.hardware, which has been obsolete for
half a year, and comp.ibm.pc.hardware, which doesn't exist on any site I
have access to, as well as to biz.comp.hardware,
Scott Purl <sdpcart@selway.umt.edu> repeated this time-worn
canard:
>If the memory speeds are different (70ns and 80ns, for example), then
>you're asking for trouble.  Get the manual for the motherboard and
>install the speed memory that it specifies.  Mixing speeds leads to
>memory parity errors, which leads to all sorts of wierd things
>(spontaneous reboots and hangs, for example).

Email to sdpcart@selway.umt.edu went unanswered.

There is no such thing as "matched" SIMMs.  Identically marked
SIMMs can be 20 ns different, and are likely to be at least 5 ns
different.  SIMMs marked a grade apart can be within a nanosecond
of each other.  That's because DRAMs are *always* faster than
they are marked.  They can be *much* faster if, for example,
the manufacturer's process is turning out nothing but 60 ns parts
but customers are ordering mostly 80 ns parts.
(This was explained to me by the man who manages the largest DRAM
plant in the world, who was visiting to try to sell me ASICs made
in the same plant.  You may have "heard" a different story from the
guy in the corner clone store.)

As I've said before, I believe what is really going on here is
a big scam.  Dealers are selling motherboards without testing them
properly.  Many of these low-cost boards cut corners and do not
meet the *highly standardized* requirements of DRAM SIMMs.
Their strobe transitions aren't monotonic.  The strobes ring.
The supply rails are noisy.  Pulsewidths aren't met.
Access times are cheated.  (I have interviewed with two
motherboard companies where I was asked which DRAM specs can
be cheated and which must be observed.  Scum!  Schlockers!)

These boards will usually work when the DRAM array is not fully
stuffed.  The problems begin when the user adds more RAM,
because that's when the signal integrity and time cheats show up.
Such a dog board often can be made to work by "cherrypicking" DRAMs,
that is by trying a few pieces until a "golden" set can be found
that can tolerate that particular motherboard's problems.

The dealer then faces a problem, how to get the customer to
pick cherries at the customer's expense.
Tell them "these SIMMs aren't matched, you need a matched set"
is often the answer.

This has been going on for so long that it's creeping into
OS documentation.  At least one commercial OS product prints
"mismatched RAMs" when its memory initialization fails.

A digital system well-designed for volume production will always
accept *faster* chips than were specified.  You can "mix" them in
the same "bank" or between "banks".
The exception to this rule is for motherboards which sense the
grade of 72-lead SIMMs and automatically configure "wait states" or
other timing behavior: these should have the slowest parts in
the socket that does the sensing.

Don't buy *anything* from a dealer who tells you stories about
"matched" SIMMs.  Ask them to show you oscilloscope traces.

>Scott Purl
>Calamity Factors -- Terra Simulations and Computing
>sdpcart@selway.umt.edu

I'd like to know *what* Scott simulates.  I hope it's nothing electronic.
Scott's advice to get the motherboard manual is good, though.

Cameron in San Jose, speaking only for myself.


-- 
Cameron Spitzer, San Jose CA   cls@truffula.sj.ca.us