🚀 go-pugleaf

I suspect that these are for programs that don't break cache too much. With
the ARM2, the memory system keeps up reasonably well. With the ARM3 the cache
helps to hide the effects of slow memory, though as soon as the cache breaks
the performance drops to that of an ARM2, or slightly worse. The ARM600 gains
a write back buffer which decouples the core from memory even more, but again
if you have massive inline code you can slow the processor right down.
       Rabin Ezra
-- 
Rabin Ezra, PhD Student                         Internet: rabin@dcs.qmw.ac.uk
Dept of Computer Science,                       UUCP:  rabin@qmw-dcs.UUCP
Queen Mary and Westfield College,               JANET: rabin@uk.ac.qmw.dcs
Mile End Road, London E1 4NS, G.B.

cs89jmb@brunel.ac.uk (Aeon Flux) writes:


>Salutations,

>Not being a real techy I was pondering this question last night and wondering
>why I couldn't come up with an answer.

>An Arm2 runs at 8Mhz and has approx "mippage" of 4
>An Arm3 runs at 26Mhz and has approx "mippage" of 13

>Rumour has it that the next ArmX chip runs at 50Mhz and has approx mippage of
>28

>Now then, what I don't understand is the Arm3 has a 4k cache which helps to 
>speed things up, right?
>The Arm3 (most common ones) are approx 3 times faster than an Arm2 - is this 
>because they are clocked 3 times faster?

Basically, yes.

>If so, where does the cache help?

If you run an 26MHz ARM3 from 8MHz memory, it can't get its
instructions fast enough to run at the full 26MHz. If the instructions
happen to be in the cache (which runs at the same speed as the
processor) it can run these at full speed. The cache also helps on
loads, if the value to be loaded happens to be in the cache. On the
ARM3 the cache doesn't help on stores - these are still written to
main memory. On the ARM600/ARM610 a write buffer makes the processor
able to continue with the next instruction before a store is
completed.

On the A5000 and A4 the main memory is 12MHz, which helps on memory
accesses that doesn't use the cache. Thus the A5000 is faster than an
A300/A400 with ARM3 added, though the difference is not so great as
between ARM2s running from 8 or 12MHz memory (like the new A3010
etc.).

I expect the ARM700 (or whatever) to have a write buffer similar to
the A600, and I expect it to run at better than 30MHz. 50MHz is
possible - it mainly comes down to fabrication techology, and some of
the current manufacturers of ARM chips should have the necessary
technology. It may make the chips more expensive, so maybe only the
high end machines will run at this speed.

A more important question is what MEMC it is going to use. The MEMC1a
is too limited in the amount of memory at can address. The ARM600 has
a built-in MEMC that is capable of addressing all the memory you are
likely to want to put on an ARM based computer, but it also has
several other advanced features which assists protection and garbage
collection, features which RISC OS is not likely to be able to use,
and RISCix neither. So maybe we will see a MEMC with a similar
translation table structure as on the ARM600, but with a simpler
protection scheme (similar to that of MEMC1a).

A machine using 50MHz ARM700, a good MEMC, VIDC20, an FPU (which
should be just about ready) and simple memory extension up to or above
128MB will be quite impressive, but also somewhat more expensive than
even the A540 is today.

	Torben Mogensen (torbenm@diku.dk)