[H-GEN] Tag RAM and lots of memory

[David Jericho]

-----BEGIN PGP SIGNED MESSAGE-----

On Thu, 25 Sep 1997, Matthew Tippett beat me to the punch and wrote:

> Martin Pool wrote:
> > 
> > I read something a while ago which more or less said:
> > 
> >    If you put more than 64MB in a PC, you have to buy more tag
> >    RAM/cache/something else or it will slow right down.  PCs
> >    can't normally cache the memory above 64MB.
> > 
> > Perhaps somebody on the list with a studly machine or a bit of general
> > knowledge can enlighten me?
> > 
> > Martin Pool

> How much cache you need for a given amount of memory depends
> on your MB manufacturer.  Most that I have seen let 256k
> support up to 64 MB.  

The amount of cache in a machine really makes no difference to how much
ram you can put in a machine. As Matt explains later on, the TAG cache is
what determines how much ram you can "sensibly" put in.

> WR2 TAG RAM memory, that typically is associated with the amount
> of cache you have.  (IIRC it is used to say what is dirty and what
> is clean).  If you get cache modules you will get it as a all in
> one package (as in ASUS boards for example).  If you need to get
> individual chips then follow your motherboard manual's directions.

Very few motherboards now days have user installable cache. Only
those that are of a reputable brand have the option and only on the
Pentium boards. PPros have on chip cache, and IIRC on chip tag cache.

> When you have more memory than the cache architecture is designed to
> handle you start to get misses (read the system gets the data from the
> memory, not the cache).  This has a similar effect to have little or no
> cache.  Add more cache memory and then the misses go down and things go
> back to normal.

Hence a machine that I put together for a l^Huser recently. The user
couldn't understand why their new super duper P233MMX machine with its el
cheapo motherboard was slower than their old P133 under Windows.

It turns out that the motherboard could only cache 64MB (IIRC, TX chipsets
can't cache more) and low and behold, they had 128MB. AFAIK, Windows 95
fills top down, so guess which lot of RAM wasn't being cached. It was
painfully slow when doing things like loading word.

Problem was fixed by convincing them they needed a higher quality
motherboard. I did my Micronics plug :)

> So in reality it doesn't slow things down, it just doesn't
> work efficiently.  

I'll second that. If anyone wants a demonstration, I'll fire up my PPro at
the next meet with the external cache disabled. 2.07 bogomips in total
over two cpus... cranking dude :) Cache makes little difference on
non-pipelined non-superscalar cpus, but without cache, a PPro, Alpha or
the likes is an ugly dog.

As I tell people, if you are planning on getting a study machine, make
sure you buy a motherboard to match. A mismatched chip/motherboard is just
downright wasted money.

- --
- -- David Jericho - paper trails and me just doesn't work
- -- Finger s343974 at student.uq.edu.au for PGP key
- -- <META HTTP-EQUIV = 'REFRESH' CONTENT = '0; http://www.microsoft.com/'>

-----BEGIN PGP SIGNATURE-----
Version: 2.6.3i
Charset: noconv

iQCVAwUBNCohYL4XfRuE8i8NAQHXqQP8CLxxh2nVnlJJ/dsL2llWvpRbFemG7kqb
yjYr6BnOQyhyUnCzDu9uNUsE9MjEJXy6+XZwnxBUnl9zWZ8U0WuuAJ+jSegXVKbm
/Hep23ZdFZaIn4fxJy06HnfFg0Kvs9FC+CTClsHOdNDI55HsDvWkMvr9/fasNQ25
ZJpAqKZz4Kw=
=3sCl
-----END PGP SIGNATURE-----

----------------------- HUMBUG General List --------------------------------
echo "unsubscribe general" | mail majordomo at humbug.org.au # To Unsubscribe