Y Combinator is restoring one of Alan Kay's Xerox Alto machines

[Warner Losh]

On Fri, Sep 9, 2016 at 2:56 AM, Eric Smith <spacewar at gmail.com> wrote:
> Note that the speed grade of a DRAM only guarantees that it is at
> least as fast as the grade; it may be faster.  For instance, an
> MK4116-2 is rated for maximum 150 ns access, while an MK4116-3 is
> rated for maximum 200 ns access.  If a machine is designed to work
> with the MK4116-3 (maximum 200 ns access), either the MK4116-2 or
> MK4116-3 will work, and the machine cannot distinguish them, as it's
> entirely possible and likely that many 4116-3 chips are under typical
> conditions actually just as fast as the 4116-2.  This is true whether
> the mix of -2 and -3 parts are in the same bank or different banks.

I once worked on an ISA board that was perfect. The driver worked, we
got good data. We could control the system we needed to based on a
stream of measurements from the board. We moved on to other projects,
but maybe 6 or 7 years later another customer had a need almost
identical to the first. We found we didn't have enough parts left over
from before so we had to buy new FIFOs to build new boards. The FIFOs
we used were 200ns parts (I forget the number), so we bought a bunch
and set to building new boards. In test we found the card was
misbehaving once in a while (like maybe once a day). Of course, the
driver was blamed (I hated that, as the driver writer). Long story
short: the hardware design had some asynchronous elements that
depended on the 200ns timing. The FIFOs were actually a faster 100ns
version that was marked for 200ns because the times were a maximum and
it was easier to make one speed grade and sell it into many different
legacy markets. Once we corrected the asynchronous behavior in the
circuit, everything worked again.

So even getting an identical part number won't guarantee an identical
part due to changes in process :(

Warner