Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!cbmvax!daveh
From: daveh@cbmvax.commodore.com (Dave Haynie)
Newsgroups: comp.sys.amiga.hardware
Subject: Re: Hurricane 2800 doesn't like 16 bit fast ram?
Message-ID: <14917@cbmvax.commodore.com>
Date: 5 Oct 90 21:01:23 GMT
References: <4825@crash.cts.com>
Reply-To: daveh@cbmvax.commodore.com (Dave Haynie)
Organization: Commodore, West Chester, PA
Lines: 69

In article <4825@crash.cts.com> lkoop@pnet01.cts.com (Lamonte Koop) writes:
>koren@hpfelg.HP.COM (Steve Koren) writes:

>>Someplace in the Hurricane 2800 docs, they mention a "bug in AmigaDos"
>>which they claim prevents '030 data and instruction caches from working
>>correctly with normal fast ram.  

I read this when I wrote a review of the Hurricane boards.  It's total 
nonsense.  Basically, the rules for data caching hadn't been invented at the
time the 68000 bus, and by extention the Zorro II bus, were defined.  So, in
order to support data caching of 16 bit RAM, your 68030 board needs extra
logic.  The Hurricane board didn't add this extra logic, the A2630 and
GV 68030 boards did.  I don't hold this against the Imtronics, the current
vendor of the Hurricane boards; they bought the Hurricane design from the
inital designers.  Both solutions work, the only real problem I see is that,
when the documentation says it's a bug in the Amiga architecture that prevents
data caching of 16 bit memory, it is telling you a baldfaced lie.

>The data cache is not enabled on the 16-bit RAM, but there is a PAL to do this
>if I want it (never saw the need...16-bit RAM cacheing probably won't add much
>perfomance).

On the contrary!  The greater the difference between cache speed and memory
speed, the more the cache helps.  If the memory were as fast as the cache,
the cache would offer no speed improvement in a single processor system.

>I'm not sure about this (someone correct me if I'm wrong), but I believe the
>instruction cache should work fine in the 16-bit environment...the OS itself
>will enable the ICache, and there are boards out there which have no 32-bit
>RAM, but run 020s and 030s with the ICache on.

That's true.  AmigaOS 1.3 knows about 68020s, but not 68030s.  So the I-Cache
is automatically enabled, but not the D-Cache.  The hardware restriction for 
data cachability on the 68030 never applied to instruction cache, since that
was a read-only cache and instruction sizes were well defined.  The D-Cache
requires hardware reads to always return the full port size of the addressed
item, regardless of the instruction-requested data object size.  This wasn't
natural for a 68000->68020 CPU, where it made sense to make read and write
actual sizes the same.  It's trivial to design a 68030 Coprocessor board that
will always request full port reads from 16 bit memory to allow cachability, 
but the folk at Ronin, who did the Hurricane design, decided to sacrifice a
little performance and allow only D-caching of 32 bit memory.

>In fact, I've done some testing about, and found that in a 16-bit RAM 
>environment, it's possible to make a 25MHz 030 run SLOWER than a 7.14MHz 68000, 
>IF the Instruction cache is off, and no 32-bit RAM is used.  The 020, 030, etc... 
>are all very reliant on the ICache as in integral part of their architecture,
>and to not use them in a 16-bit environment really exposes the processor to 
>the bottleneck of said environment.  

The basic problem here is that the instruction unit in both 020 and 030 always
pre-fetches full 32 bit words.  If you have 16 bit memory, that's two bus
cycles.  But, in many cases, the second 16 bit word is thrown out.  This is
never a penalty on a 32 bit bus, but on a 16 bit bus, you end up fetching
and throwing out stuff.  A 68000 only does 16 bit prefetch, and therefore never
throws out anything in the instruction stream.

>In short, I'd be just a bit suspicious about the claim that a bug in AmigaDOS
>prevents the ICache from woring properly in a 16-bit environment.  But I COULD 
>be wrong....

But you aren't, and you obviously had the good sense to spot a bad argument.

>UUCP: {hplabs!hp-sdd ucsd nosc}!crash!pnet01!lkoop

-- 
Dave Haynie Commodore-Amiga (Amiga 3000) "The Crew That Never Rests"
   {uunet|pyramid|rutgers}!cbmvax!daveh      PLINK: hazy     BIX: hazy
	Standing on the shoulders of giants leaves me cold	-REM