Re: Memory access

From: Johan Klockars <rand_at_cd.chalmers.se>
Date: Tue, 13 Jan 1998 14:57:36 +0100 (MET)

> Well it's GEMBENCH..... I know that it's not the best to use but it's strange
> that it should be so expeptinal wrong as it sounds.

No, that's not strange at all. GEMBench is an incredibly stupid benchmark.

It's been a long time since I looked at the memory tests, but I seem to
recall them being made in a way that was very cache unfriendly on an '030,
but rather nice to the '040 caches.
I don't know how the '060 caches are organized.

> the roof.... they were at 2400%. Sounds more correct. Maybe the AB040 fools

It does?

A normal Falcon has something like 6Mbyte/s bandwidth to RAM. 24 times
that would be 144 Mbyte/s, which given a 32 bit bus would require all
accesses to take less than a single clock cycle at 32MHz.

Let's say that it was really 128Mbyte/s to keep within the physical limits.
That still requires an instant setup of each burst transfer, which just
isn't possible. I'd be very surprised if even 2-1-1-1 was possible.

An AB040 can reach slightly more than 30MByte/s and I'd say that the
Hades can at best manage a little more than double that. I'd be surprised
if it did, though.

> But also, Petr did say that Quake on the Hades were slower than on the AB040 so
> something fishy is going on in the Hades.

An '060 should be _very_ good at Quake.
I don't think the difference between a Hades040 and an AB040 should be
significant, though.

-- 
  Chalmers University   | Why are these |  e-mail:   rand_at_cd.chalmers.se
     of Technology      |  .signatures  |            johan_at_rand.thn.htu.se
                        | so hard to do |  WWW/ftp:  rand.thn.htu.se
   Gothenburg, Sweden   |     well?     |            (MGIFv5, QLem, BAD MOOD)
Received on ti. jan. 13 1998 - 20:41:00 CET

This archive was generated by hypermail 2.3.0 : ti. nov. 03 2015 - 20:07:53 CET