Tell me about CBM A3640 accellertor please!
D. Colic
I have seen add for CBM A3640 040 accellerator. Could someone give me
some specs. egs memory, type speed etc. thanks.
J_PL...@amtrash.comlink.de
(19.04.1995)
DC> I have seen add for CBM A3640 040 accellerator. Could someone give me
DC> some specs. egs memory, type speed etc. thanks.
DC>
Lowcostboard (not very fast)
~18 MIPS
uses mainboardmemory (no StaticColoumn mode->exchange first ZIP in A3000)
->memoryspeed less than 50% of a 030-A3000/A4000
latest version (3.2)is not able to use 60ns ram (skipmode)
noL2cache
useful in A3000(T)/4000
memoryspeed less than 50% of a 030-A3000/A4000
no burstmode
reduced Z3 speed (is any 040 able to burst Z3max up to 150MB/s ?)
Kick2.04+ is required >no Bootroms!<
Use version 3.2 if there are problems...
runs synchronus to the mainboard (25 MHz..not more not less)
Maprom support (kickstart in ram without using the MMU..maped via "PAL"-hardware
Be sure that it works in YOUR machine before buying. There are many people out there
with an A3640 lying on their table...not placed into the A3000.....
I would select Warpengine or Cyberstorm for my machine
Made my Merlin work... (the original 030 fails)
/ PLEW...@informatik.fh-hamburg.de \
/ J_PL...@Amtrash.comlink.de (best!!) \
+------------------------------------------+
\ A3000T/040/18/1G Scanjet-Merlin CDTV /
\ A3000/030/6/52 StylusC-HP4L.. CD32 /
Michael van Elst
>->memoryspeed less than 50% of a 030-A3000/A4000
Incorrect.
>latest version (3.2)is not able to use 60ns ram (skipmode)
Of course, the skipmode just works by change (it strips a waitstate
for 16MHz operation and is not supposed to work at 25MHz).
>reduced Z3 speed (is any 040 able to burst Z3max up to 150MB/s ?)
The Zorro bus cannot do 150MB/s. With the current Buster you might
be able to do 50MB/s with multi transfer cycles. Single cycles
are signficantly slower (more like 10..15MB/s).
>Made my Merlin work... (the original 030 fails)
Sounds like a nice piece of hardware, that Merlin.
Regards,
--
Michael van Elst
Internet: mle...@serpens.rhein.de
"A potential Snark may lurk in every tree."
Hugh D. Gamble
No, he's right. Pick your favourite ( cache defeating ) benchmark, or
just run a real application ( e.g. emacs ).
: >latest version (3.2)is not able to use 60ns ram (skipmode)
: Of course, the skipmode just works by change (it strips a waitstate
: for 16MHz operation and is not supposed to work at 25MHz).
Skipmode was never officially supported at any speed. It works fine on
some 3000s at 25 MHz, especially with less than 16M fast RAM.
: >reduced Z3 speed (is any 040 able to burst Z3max up to 150MB/s ?)
: The Zorro bus cannot do 150MB/s. With the current Buster you might
No existing implementation can go as fast as the maximum for the
design.
--
# Hugh D. Gamble, NTT Systems Inc. (at DCIEM (416) 635-2156) Std. Disclaimers #
# hu...@dciem.dciem.dnd.ca, hu...@kink.PhaedraV.On.Ca #
# "automata da vida, honey" - CA Butterfly #
Michael van Elst
>Michael van Elst (mle...@serpens.rhein.de) wrote:
>: J_PL...@amtrash.comlink.de writes:
>: >->memoryspeed less than 50% of a 030-A3000/A4000
>: Incorrect.
>No, he's right. Pick your favourite ( cache defeating ) benchmark,
I do. I wrote such a benchmark. It is fairly accurate. It shows
the effect of a copyback cache.
>just run a real application ( e.g. emacs ).
That's significantly faster on the A4000/040. What's your point ?
>Skipmode was never officially supported at any speed.
It is basically the same as the external 16/25MHz jumper of the
older Ramsey. It eliminates one waitstate to get a 4clock cycle
instead of a 5clock cycle.
>No existing implementation can go as fast as the maximum for the
>design.
But they usually come close :)
Hugh D. Gamble
^^^^^^^^^^^^^^^^^^^
: I do. I wrote such a benchmark. It is fairly accurate. It shows
: the effect of a copyback cache.
There's a good market for rigged demos, they make people happy. :-)
The caches on the '040 are only 4k, that doesn't take you far unless
you are doing hand crafted floating point or something else with
relatively compact memory requirements.
: >just run a real application ( e.g. emacs ).
: That's significantly faster on the A4000/040. What's your point ?
Doing what? I've benchmarked long emacs operations at half the speed
on a 3000T w 3640 as on the same machine with the motherboard '030.
About exactly the difference in the memory speed. Get out your
stopwatch and try hanoi 16 sometime, with and without the '040.
: >Skipmode was never officially supported at any speed.
: It is basically the same as the external 16/25MHz jumper of the
: older Ramsey. It eliminates one waitstate to get a 4clock cycle
: instead of a 5clock cycle.
No, skipmode isn't a straight wait state reduction, but it has a
similar effect sometimes. I probably have Dave Haynie's description of
it archived somewhere, I can't do it justice.
: >No existing implementation can go as fast as the maximum for the
: >design.
: But they usually come close :)
No. Like you said in your post, none of them come close to 150MB/s.
Dennis Grant
(Hugh D. Gamble) wrote:
> Doing what? I've benchmarked long emacs operations at half the speed
> on a 3000T w 3640 as on the same machine with the motherboard '030.
> About exactly the difference in the memory speed. Get out your
> stopwatch and try hanoi 16 sometime, with and without the '040.
So is this problem with the 68040 card?
I have a 4000/030, and I'm getting a 3640R3.1 to speed up my render times.
I currently have 60ns RAM on all my SIMMS, and I am using SpeedRamsey to
get the 4 cycle mem access.
So, when the new card arrives,
1) Will my mem access slow down?
2) Can I still use SpeedRamsey?
3) Should I expect a net speedup or slowdown in
a) render speed b) screen redraws
> # "automata da vida, honey" - CA
Butterfly #
---------------------------------------------------------------------
Dennis Grant Cycor Tech (Amiga&Mac) Support / HTML & Graphic Design
dgr...@cycor.ca Amiga 4000/040/6/970/17"IDEK/2XCD-ROM/14.4 AmiTCP
http://www.cycor.ca/TCave/ Visit Trog's Cave!
Hugh D. Gamble
: > Doing what? I've benchmarked long emacs operations at half the speed
: > on a 3000T w 3640 as on the same machine with the motherboard '030.
: > About exactly the difference in the memory speed. Get out your
: > stopwatch and try hanoi 16 sometime, with and without the '040.
: So is this problem with the 68040 card?
The A3000 Motherboard, and the Zorro bus interface were designed as a
good match for the '030. It is not possible to get fast '040 memory
access through the CPU slot, so in that sense it's not the fault of
the '040 card. The only workaround is to put RAM locally on the CPU
card.
: 1) Will my mem access slow down?
For large data accesses it will be about half as fast, but rendering
code is more FPU bound than RAM bound. You should see a big speedup (
4-10 times? ) if you have rendering s/w optimised for the '040 FPU
subset.
: 2) Can I still use SpeedRamsey?
Perhaps. P.S. How much fast ram do you have? More than 8M?
: 3) Should I expect a net speedup or slowdown in
: a) render speed b) screen redraws
Rendering will many times faster, screen redraws may be a tad faster.
--
# Hugh D. Gamble, NTT Systems Inc. (at DCIEM (416) 635-2156) Std. Disclaimers #
# hu...@dciem.dciem.dnd.ca, hu...@kink.PhaedraV.On.Ca #
Hugh D. Gamble
: can access motherboard memory about as fast as the 68030. With
No. If you don't believe the empirical performance numbers, you can go
to the engineering specs and add up the wait states and cycle times.
It still takes twice as long to get words to an '040 through the CPU
slot as it does to get them to the '030. Now, reading them from the
cache inside the '040 chip is faster than reading them from the cache
inside the '030, and the cache is a little bigger, but that's a
different issue.
: common benchmarks like AIBB you will see a significant drop
: in speed of memory copy operations. This is caused by the 68040
: copyback cache that effectively _reads_ every word before it is
: overwritten in the cache. That's why Motorola invented the MOVE16
Try this: disable copyback cacheing and run it again. Compare with
AIBB numbers for the motherboard '030.
Hugh D. Gamble
: >good match for the '030. It is not possible to get fast '040 memory
: >access through the CPU slot,
: Of course it _IS_ possible, it just costs more. The A3640 is a cheap
: design and an old design.
It should be possible to do better than the 3640. A CPU card could
probably add support for burst mode through the CPU slot, for e.g. and
get the averaged access time down. Memory access would still be
slower than the '030, just not twice as slow.
Hugh D. Gamble
: Booted fine - which was a suprise - as I had installed a Maxtor 850Mb IDE
: hard drive slaved to my old Seagate 120 the week before, and that had
: resulted in a system that would not cold boot. (You had to warm boot the
: machine on powerup)
It takes longer to boot the 3640 and your drive probably has enough
time to spin up first now.
: Anyway, playing around with the machine since then has proved that it is
: now MUCH faster. Anything that was calculation-intensive has showed a
Happy happy, joy joy. You can render more Yak shavings than ever before :-)
: SysInfo (ick!) is reporting a 4 times speed increase, and, interestingly,
: a _decrease_ in IDE transfer rates of about 400k/sec (although this could
: just be faulty calculations on SysInfo's part)
No, IDE depends heavily on your CPU for data transfer and you're
seeing the effect of the slower memory access. A good SCSI card will
minimise the problem by doing DMA without going through the CPU. It's
smart not to trust benchmarks though. You can time your disk access
with and without the 3640 with a stopwatch by hand.
Michael van Elst
>The A3000 Motherboard, and the Zorro bus interface were designed as a
>good match for the '030. It is not possible to get fast '040 memory
>access through the CPU slot,
Of course it _IS_ possible, it just costs more. The A3640 is a cheap
design and an old design.
Of course you will never get faster access to motherboard memory than
what Ramsey can do, that's a Ramsey limit and not an interface limit.
Local RAM on the card can be faster.
>: 1) Will my mem access slow down?
>For large data accesses it will be about half as fast,
Data _reads_ are about same speed for 68040 and 68030.
Michael van Elst
>It should be possible to do better than the 3640. A CPU card could
>probably add support for burst mode through the CPU slot, for e.g. and
>get the averaged access time down. Memory access would still be
>slower than the '030, just not twice as slow.
Memory access would be _exactly_ the same as the 68030. Even the
A3640 isn't really slower ! A good design could hit memory faster
than the 68030.
Bert Heise
: >It should be possible to do better than the 3640. A CPU card could
: >probably add support for burst mode through the CPU slot, for e.g. and
: >get the averaged access time down. Memory access would still be
: >slower than the '030, just not twice as slow.
: Memory access would be _exactly_ the same as the 68030. Even the
: A3640 isn't really slower ! A good design could hit memory faster
: than the 68030.
As did the Cyberstorm40 in my A3000T, before I purchased local Cyber-Ram.
The Static Ram on the A3000 motherboard was slightly _faster_ than with
the original 030.
: Regards,
: --
: Michael van Elst
Bert.
Michael van Elst
>No. If you don't believe the empirical performance numbers, you can go
>to the engineering specs and add up the wait states and cycle times.
>It still takes twice as long to get words to an '040 through the CPU
>slot as it does to get them to the '030.
Then you don't believe measurements ?
>Try this: disable copyback cacheing and run it again. Compare with
>AIBB numbers for the motherboard '030.
I did something different and tested read and write speeds independently
(AIBB measures a copy operation).
Neither the 030 nor the 040 reached theoretical speed (200ns cycles)
for reading memory. Both do about 13MByte/s.
Neither the 030 nor the 040 reached theoretical speed for writing
memory. The 68030 moved about 17MByte/s, the 68040 moved about
14MByte/s without copyback and just half of it with copyback.
Regards,
--
Michael van Elst
Internet: mle...@serpens.rhein.de
Michael van Elst
>As did the Cyberstorm40 in my A3000T, before I purchased local Cyber-Ram.
>The Static Ram on the A3000 motherboard was slightly _faster_ than with
>the original 030.
Correct :) Same for the 68060.
Hugh D. Gamble
: The Static Ram on the A3000 motherboard was slightly _faster_ than with
The A3000 is not capable of accepting static RAM. You can use DRAM
with something called static column mode, which allows you to only
change the low address bits for faster localised access after the
first one.
Bert Heise
: Bert Heise (be...@news.hrz.uni-dortmund.de) wrote:
: : The Static Ram on the A3000 motherboard was slightly _faster_ than with
: The A3000 is not capable of accepting static RAM. You can use DRAM
: with something called static column mode, which allows you to only
: change the low address bits for faster localised access after the
: first one.
Booohoooo - of COURSE I meant Static Column RAM - I only was a bit lazy writing
the whole thing, but I thought everybody (with a few exceptions... ;-) ) would
know what I'm talking about... (also in use: SCRAM)
: # Hugh D. Gamble, NTT Systems Inc. (at DCIEM (416) 635-2156) Std. Disclaimers #
: # hu...@dciem.dciem.dnd.ca, hu...@kink.PhaedraV.On.Ca
Bert.
J_PL...@amtrash.comlink.de
(25.04.1995)
ME> J_PL...@amtrash.comlink.de writes:
ME>
ME> >->memoryspeed less than 50% of a 030-A3000/A4000
ME>
ME> Incorrect.
Please don`t simply write "Incorrect!"... A3640 isn`t able to burst, static-column fails...
busspeed for example reports a lower transfer...WHY???
ME> >latest version (3.2)is not able to use 60ns ram (skipmode)
ME>
ME> Of course, the skipmode just works by change (it strips a waitstate
ME> for 16MHz operation and is not supposed to work at 25MHz).
But it has nothing to do with A3640`s PalUpdate?!
My A3640-3.0 did the skipmode in the same machine (maybe another CPUBoard?).
I am using 60ns-RAM and the skipmode_option WAS designed to use 60ns Ram.
Commodore didn`t plan to sell a A4000-16Mhz?!
Haynie asked the man who did Ramsey-07...it should work, but it doesn`t...bäääh.
ME> >reduced Z3 speed (is any 040 able to burst Z3max up to 150MB/s ?)
ME>
ME> The Zorro bus cannot do 150MB/s. With the current Buster you might
ME> be able to do 50MB/s with multi transfer cycles. Single cycles
ME> are signficantly slower (more like 10..15MB/s).
Are you able to do 150MB/s with any PCI-Chipset currently used?
ME> Sounds like a nice piece of hardware, that Merlin.
Nice? Hmmm....
Michael van Elst
>Please don`t simply write "Incorrect!"... A3640 isn`t able to burst, static-column fails...
>busspeed for example reports a lower transfer...WHY???
A3640 isn't able to burst, right. But that's not really important,
especially when you test the memory with data accesses.
>But it has nothing to do with A3640`s PalUpdate?!
No.
>My A3640-3.0 did the skipmode in the same machine (maybe another CPUBoard?).
>I am using 60ns-RAM and the skipmode_option WAS designed to use 60ns Ram.
No. It was designed for 16MHz operation.
>Are you able to do 150MB/s with any PCI-Chipset currently used?
No, but about 100MByte/s is possible.
J_PL...@amtrash.comlink.de
ME> >Please don`t simply write "Incorrect!"... A3640 isn`t able to burst, static-column fails...
ME> >busspeed for example reports a lower transfer...WHY???
ME>
ME> A3640 isn't able to burst, right. But that's not really important,
ME> especially when you test the memory with data accesses.
Thanks!
ME> >My A3640-3.0 did the skipmode in the same machine (maybe another CPUBoard?).
ME> >I am using 60ns-RAM and the skipmode_option WAS designed to use 60ns Ram.
ME>
ME> No. It was designed for 16MHz operation.
The old A3000 versions were designed for 16MHz
(there was a jumper on the mainboard etc.).
The -07 still recognices the jumper setting...to be compatible.
1M*1 mode register was removed and the skip-register was implemented.
It should work with A3640 and it does if you use 60ns ram.
ME> >Are you able to do 150MB/s with any PCI-Chipset currently used?
ME>
ME> No, but about 100MByte/s is possible.
Is there any PIC availabel which supports rates like 100MB/s?
Michael van Elst
>The -07 still recognices the jumper setting...to be compatible.
>1M*1 mode register was removed and the skip-register was implemented.
>It should work with A3640 and it does if you use 60ns ram.
Skip-mode at 25MHz reduces cycle time by 1 clock == 40ns. How can
a 60ns chip be 40ns faster than 80ns ? At least you reduce the
safety margins by 20ns which might work.. or not. Some people
used 60ns chips for some time and reported problems during hot
days.. a clear sign that 60ns is not fast enough to meet specs.
It might work for you though.
>ME> No, but about 100MByte/s is possible.
>Is there any PIC availabel which supports rates like 100MB/s?
Probably not. But a shared bus needs a higher bandwidth than
required by a PIC anyway.
Dave Haynie
>J_PL...@amtrash.comlink.de writes:
>>The -07 still recognices the jumper setting...to be compatible.
>>1M*1 mode register was removed and the skip-register was implemented.
>>It should work with A3640 and it does if you use 60ns ram.
>Skip-mode at 25MHz reduces cycle time by 1 clock == 40ns. How can
>a 60ns chip be 40ns faster than 80ns ?
Well, it can't. However, it can be 40ns faster than 100ns, which is
RAMSEY's basic cycle requirement, same as on the A2630. The 80ns
speed requirement is for the static-column cycles (burst and
page-detect), which are unchanged in skip mode. Yes, this does imply
that 100ns page-mode DRAM will work in an A3000 or A4000 (most,
anyway; some very old 100ns DRAM may have parameters other than RAS
time that are too slow, but most will work).
>Some people used 60ns chips for some time and reported problems
>during hot days.. a clear sign that 60ns is not fast enough to meet
>specs. It might work for you though.
One problem is that 60ns DRAM aren't always 60ns DRAM. Or, basically,
that there are no real speed standards. All 60ns tells you is that RAS
access happens in 60ns. It doesn't speak to any other DRAM
parameters. In fact, from what I recall Greg stating, some 70ns DRAM
could possibly work. If C= had every officially supported skip mode,
they would have had a list of approved vendors.
Dave Haynie | ex-Commodore Engineering | See my first film
Sr. Systems Engineer | Class of '94 | "The Deathbed Vigil"
Scala Inc., US R&D | C= Failure n. See: Greed | in...@iam.com
"Caught a bolt of lightning, cursed the day he let it go" -Pearl Jam