Conference hydra::amiga_v1

    re: .0--``...without 32 bit RAM, a 14 MHz 68020 should only perform
    as fast as a 14 MHz 68000 anyway''.
    
    I understand that a 68020 doesn't improve the Amiga's performance
    much in the absence of 32-bit RAM.  That leads to the question of
    how much this product will improve the Amiga's performance.  If
    the 7 MHz 68000 takes every available memory cycle already, speeding
    it up isn't going to improve system performance.
        John Sauter

    Doesn't going to 14 MHz double the number of cycles available?
    
    It was my understanding that going to a 68020 with 32 bit RAM would
    provide a four-fold performance increase.  2X due to the higher
    clock rate, and another 2X due to the doubling of the size of the
    data path.
    
    In reality, just going to the 68020 produces almost no speed increase
    because, I am told, that the '020 must perform extra instructions
    to fetch 16 bit wide data.  This performance hit is so bad that
    the system approaches a vanilla 7 MHz 68000.
    
    The Hurricane accelerator board produces a genuine speed increase
    without 32 bit RAM.  How it does this is a mystery to me.
    
    Perhaps Randy Meyers can jump in here and give us a hand...
    
    Ed.
    

>        In reality, just going to the 68020 produces almost no speed increase
>    because, I am told, that the '020 must perform extra instructions
>    to fetch 16 bit wide data.  This performance hit is so bad that
>    the system approaches a vanilla 7 MHz 68000.

    This is a war story floating around following the _Amigaworld_ review
    of the CSA 68020 accelerator for the A2000.  The reviewer ran AmigaDOS
    v1.1, which does not turn on the '020 instruction cache.  Without
    the cache on the processor cannot fill all the bus wait states doing
    instruction execution.  Do you really trust _Amigaworld_? :-)    

>    The Hurricane accelerator board produces a genuine speed increase
>    without 32 bit RAM.  How it does this is a mystery to me.

    No mystery.  The CSA board has the same performance improvement when
    run with the current version of the OS.  I have a Motorola application
    note which shows performance numbers increasing 50% on average by
    popping in a 68020 adapter board to one of their evaluation modules
    (same clock speed but several wait states). Send email for a copy.
    
    Once false information starts circulating, it's soooo hard to stamp
    out.
    
    Wes
    (Hardware type) 

    Maybe my reply in -.1 was a little abrupt.  Here is where I think
    the performance improvement or hit comes from in a 68020 board running
    in a stock Amiga.  The first column is my rating of the effect,
    followed by the feature itself.
    
    +	Higher clock speed
    +	Cache memory (can be canceled by not running v1.2)
    +	Fewer cycles per instruction
    0	New '020-only instruction modes
    -	Instruction fetches always 32 bits (a big + with 32-bit memory)
    
    In a standard Amiga the performance improvement comes from keeping
    the instruction pipe full to take advantage of the higher effective
    instruction rate.  Cache and 32-bit memory both make a really big
    difference.  Code with a lot of branches causes a big performance
    hit.
    
    It doesn't take more instructions to address 16-bit memory, just
    more instruction _fetches_.  This is the dark side of the 68020,
    but it is easy to overcome with the on-board instruction cache.
    
    According to a CSA marketing guy at the Motorola show in Boston
    a few months back, the _Amigaworld_ reviewer ran AmigaDOS v1.1,
    which doesn't turn on the 68020 cache.  (V1.2 detects the processor
    type and enables the '020 cache.)  If you reread the article, you
    will not see the word 'cache' even once.  This is the only 68020
    article I have ever seen without it.

    So now there is this dumb rumor floating around that a 68020 is
    worthless without wide memory, and CSA has to run ads that say,
    yes, our board really DOES improve your performance.

    Pity that the largest U.S. Amiga magazine is the direct spiritual
    descendent of that bastion of accuracy and objectivity, _Infoworld_.
    
    Wes

    
    So, what's the verdict on a 14 MHz 68000?  Is it worth a damn?
    
    Ed.
    

    >  So, what's the verdict on a 14 MHz 68000?  Is it worth a damn?
    
    There should be a modest improvement.  The "internal" cycles of
    some instructions (i.e. no bus activity going on) should take only
    half as long (on the 7 MHz bus), but the 68000 data sheet and user
    manual both appear to say that wait states must be added to instruction
    times.  In other words, the internal cycles do not overlap wait
    states.
    
    So, my opinion with no actual experience, you might see a 10-20%
    performance improvement.  The improvement is much higher for multiply,
    divide and shift instructions.  Anyone actually tried this?
    
    Wes

    
    Wes, you would seem to be correct in your assesment of speed
    improvements...  I got some more info from a CMI employee who has
    an account on PLINN (BILL*NAG).
    
    The CMI Processor Accelerator will offer no speedup on 68000
    instructions that execute in up to 4 clock cycles.  Instructions
    longer than 4 cycles will begin to show an average of 20% speedup,
    with longer instructions showing the greatest improvement, up to
    50%.
    
    This still seems like a reasonable price, considering that a 68881
    socket is included (you have to provide your own crystal for the
    math chip).  If a 68010 provides a 5% boost for $50, that <$200
    seems fair for a 20% boost.  Sure is a lot cheaper than an '020.
    
    CMI has begun small shipments of the board.  I think 40 unit went
    out this week, with more coming next week.
    
    At AmiExpo, another company showed a similar product priced at $139.
    
    Ed