|
4/30/88
The Hurricane accelerator card for the Amiga 2000 is the same card used
on the Amiga 1000 and 500 with the addition of an adapter card called a
K-Card. This K-Card is a small PC board that plugs into the 86 pin
co-processor slot in the 2000 and contains only a 64 pin IC socket. The
Hurricane accelerator card plugs into this 64 pin socket and then the entire
assembly slips into the co-processor slot. This makes for a wide sandwich
but it does not interfere with any other slot. There is a 96 pin DIN
connector on the top of the card that will accept a 32 bit memory card.
To me, this socket seems very close to the 2000 case and I don't see how a
memory board can attach to the card. I'll have to discuss this with Finally
Technologies at a later date. The 32 bit dynamic memory card for the 2000
is a little ways from production so the point is moot at this time.
Installation instructions require you to remove the original 68000 chip
from its socket near the power supply and store it away in anti-static
wrapping; apparently the accelerator card can not disable the motherboard
CPU socket. Be careful when removing the old 68000 chip, it's real easy to
damage the traces on the motherboard under the chip by levering the chip out
of the socket.
The accelerator card comes with a 14Mhz 68020 chip with a 16Mhz 68881
math coprocessor chip. Higher speed 68881 chips are available at higher
cost. Several handy programs come with the board. One program patches
AmigaBasic to correct an illegal instruction that apparently the 68000
doesn't object to. Another program will scan executable programs looking
for subroutine calls to do 32 bit multiplies, divides and modulo operations
and replace them with inline 68020 instuctions. The progam can recognize
Manx and Lattice constructs; it doesn't mention any other language
compilers. Patching the executables can increase the execution of these
programs up to 15%, but once patched, they can not run on a 68000 or 68010
machine. A wedge program to trap and correctly handle the infamous Move SR
instruction is included as well as programs to turn the 020 cache on or off.
To demonstrate the power of the 68881 chip, a Mandelbrot program compiled
to use the coprocessor is provided. Benchmarks indicate that the default
image is rendered in 66 seconds on a standard Amiga 1000, 34 seconds using
the accelerator and the IEEE libraries (no 68881) and an incredable 3.1
seconds using the 68881 chip.
Also included is a version of Doctor Dobbs Savage benchmark compiled in
MAnx C. A standard Amiga 1000 executes this benchmark in 47 seconds, while
the accelerator card runs it in 1.44 seconds. One supplied version uses the
new DOS 1.3 IEEE library which can take advantage of the 68881 chip if found
or emulate the chip in software if not found. The libraries are not
included with the accelerator card. The savage benchmark if run with 32 bit
memory would execute in .82 seconds. Another savage version was compiled to
produce inline 68881 instructions; this one runs in .4 seconds without 32
bit memory and in .38 seconds with 32 bit memory.
Normal CLI activity seems to run faster by anywhere from 20 to 30%. The
time required to compile my terminal program Comm, went from 10:50
(read that 10 minutes 50 seconds) to 8:12. DirectoryMaster went from 7:34
when sorting 2150 records to 4:98. The DBW_Render program (ver 2) dropped
to 3:10 from 4:08 to generate a single row of the BackStage scene located in
the library. I am now in the process of recompiling DBW_Render version 1.0
to take advantage of the inline 68881 instructions.
I tried the same `benchmark' mentioned in the February 1988 issue of
Amiga World. Bob Ryan was testing a CSA 68020 board with 32 bit memory.
Ryan used Sculpt-3D to render the coffee cup picture with mode set to photo,
interlace, anti-aliasing set to best and size to tiny. The Amiga World
article mentioned render times of 999 seconds (16:39) without the CSA 68020
board and 256 seconds with the board. My results indicate 14:30 using a
68000 and 11:20 using the accelerator card. After using the program which
replaces subroutine calls with inline 68020 instructions, the time was
reduced to 10:40. Sculpt 3D doesn't use the 68881 chip nor the IEEE math
libraries. I am planning to buy the special Sculpt-3D version that is
compiled to use the 68881 chip. The time should be reduced significantly.
So far, in the week I've used the Hurricane board, I only found one
program that would not run. This was a sample sound player program called
CBT. The program actually ran up until the time I loaded a sound, then it
locked up tight requiring a reboot. This program doesn't work properly
on my standard 1000 either.
GOMF recognizes the presence of the 68020. GOMF2.1 claims that any of
the wedge programs that fix the Move SR instructions are not necessary, this
could be why I have no trouble with programs.
The Hurricane is available from Finally Technologies for $1000 or from
several mail order houses for ~$850.
Djj
|