[Info-vax] OT: parity vs ECC on EV4

[John Wallace]

On Oct 19, 4:12 am, Steven Schweda <sms.antin... at gmail.com> wrote:
> On Oct 17, 6:43 pm, John Wallace <johnwalla... at yahoo.co.uk> wrote:
>
> > Some AlphaStations started life with parity memory and it shipped as
> > 36bit parity (aka longword parity).
>
> > Then some HQ genius spotted that in a 21064-based AlphaStation you
> > didn't need all 36 bits. Because all memory reads and writes were
> > longwords or wider, you only needed 33bit memory to make it parity
> > protected. [...]
>
>    Really?  The oldest/lamest Alpha system I ever used was an
> AlphaStation 200 4/233.  Its approved memory was fast-page
> mode, parity.  (According to its "User Information" book,
> "The system requires 72-pin parity SIMMs that have an access
> time of 70 ns (nanoseconds) or 60 ns.")  But those extra
> "parity" bits were aggregated for use as ECC, not for simple
> parity.  One extra bit for each eight data bits means four
> extra bits for a 32-bit word, which is enough for ECC.  The
> last VMS system I saw which used parity memory as parity
> memory (and not as ECC memory) was a MicroVAX II/III or a
> MicroVAX/VAXstation 2000.  I've never seen any machine which
> used standard parity SIMM/DIMMs as anything other than ECC.
>
>    I don't get around much, so perhaps someone, somewhere,
> really did use some non-standard, parity-only,
> wider-than-8-bits, memory module, but I've never seen one,
> and I wouldn't credit such a claim without some better
> evidence than a vague recollection.

You are right to be suspicious and right to ask for better evidence.
If this short version won't do, you can try the longer version in the
21071/21072 Product Brief e.g. at
http://www.ic72.com/pdf_file/2/23605.pdf
and in the 21071/21072 Data Sheet e.g. at
http://www.compaq.com/cpq-alphaserver/technology/literature/071_72ds.pdf

The 21071 family was the intended support chipset for the EV4/21064
chip and was used in various flavours of EV4x AlphaStation and
AlphaServer boxes. The 21071 provided a 64bit wide interface to main
memory and its 21072 big brother provided a 128bit wide interface. The
64bit wide interface could use 33bit SIMMs because it didn't offer
full ECC but required parity. The 128bit wide interface in ECC mode
(such as in the AlphaServer 1000) needed more bits.

The AlphaStation 400 (and its server equivalent) was one of the
systems with a narrow bus. The EV4 AlphaServer 1000 used the wide
version.

I am reasonably confident in this picture, not least because I know of
a customer who invested in a roomful of (dozens of) AlphaStation 400s
with fully populated memory. They used to get around one crash a month
overall, which correlated reasonably with the expected memory error
rate with parity rather than with ECC, but was not what the customer
was expecting. To fix this, they had to be upgraded to AlphaServer
1000 with ECC rather than AlphaStation 400 with parity. No more
crashes to worry about.

hth
john