| I believe this is entirely a hardware question. You should ask in the
conference appropriate to the particular hardware platform:
WRKSYS::MIKASA 1000, 1000A
MVBLAB::SABLE 2000, 2000A, 2100, 2100A
MSBCS::TURBOLASER 8200, 8400
MVBLAB::ALPHASERVER_4100 4000, 4100
The interesting aspects will be
number of banks
raw latency
whether interleaving is possible
For example, the new larger memory board for Turbolaser (is that the 2GB
board?) has 4 banks instead of 2 on older ones, but is a touch slower to
access. For slot-constrained situations, it's valuable because one new
board provides as much memory (and as many banks) as two old ones. But
if you're not constrained, I think the older one is better (by a barely
measureable amount) because it's somewhat faster.
Typically interleaving is only allowed between two equal-sized chunks of
memory. Sometimes a chunk can be constructed with multiple boards.
So, for example, you might be able to interleave a 1 GB board with a
pair of 512 MB boards.
|
| What I've heard whilst kicking around the lab:
- no particular difference what you do with mikasa
- On Sable, see note 1009.1 for an interesting start - the
key variable is whether you have the 60ns or 80ns parts.
Interleaving is not a big factor.
- On Rawhide, the smaller 128mb "sync" modules are about 10%
faster than the larger "edo" modules, but this difference
will be much less with the new edo modules. Interleaving
is not a big deal. Biggest memory must come first, so if
you have mixed edo and sync the slower memory gets used first.
- On Turbolaser, interleaving is VERY important. For some tips, see:
http://tlg-www.zko.dec.com/~henning/Mem_bw.html
|