workstation speed

[Jeffrey Del Papa <dp@jasper.palladian.com>]

-------- Begin Forwarded Message
this is from a discussion on ai.works that bears on the recent debate on this list.

Resent-From: Arun <Welch%OSU-20@ohio-state.ARPA>
Resent-To: ai.works%OSU-20@ohio-state.ARPA

    Date: Wed, 29 Jul 87 08:53 EDT
    From: DON%atc.bendix.com@RELAY.CS.NET

    For those who don't know, what was formerly called LMI is now Gigamos.
    Gigamos is a Canadian holding company which purchased LMI in May
    (I think the 25th).  They say that they plan to continue development
    on what LMI was developing as well as maintaining the old line. 
    They are down to 50 employees.

the current total is around 30. (gigamos has started to hire again, so it may be more
like 40 now) LMI offically died june 11th, with the last employee (the president
handling the liquidation) staying until mid july. TI has contested the sale. The
first LMI alumni picnic is this saturday.

    Concerning the TI Explorer II, its benchmarks are 5x the explorer
    I for the processor.  Because the Explorer I spent almost all its
    time doing paging to a slow disk through a slow controller, TI has
    beefed up the controller and disk for the explorer II.  Whether this
    combination of improvements amounts to a 5x speed-up for the entire
    machine remains to be seen.

TI has improved paging performance with the latest (3.0) release, they have added a
real (temporal/ephemeral...) gc, and generally improved things considerably. No
stunning new features, they spent there time making what they had work. The compiler
is much improved, the code density is higher, and they pay a lot more attention to
optomization in the compiler. 

    While I certainly have no hard data, due to the fact that the Explorer
    I was so much slower than even the LMI, I find it hard to believe
    that the II will perform much better than Symbolics 3650.  As we
    all saw, Symbolics has announced their Lisp on a chip which they
    claim contains the microcode and several other things which TI left
    out and had to put on their processor board.  Symbolics is estimating
    a 3x speed up for upgrading 3650 and 3620's and 5x speed up for totally
    new packaging.  This speedup will definitely leave TI in the dust.

the hard numbers: we had an explorerII here for two weeks before AAAI. the simplest
raw benchmark, compiling one of our products ~400k lines of lisp, takes 90 min on a
16mb expII, 4.5 hours on a 4MW (16mb) 3620 (thus comparing apples, as they both have
the same disks) and to fuel the 68k/unix/lucid/etc.. debate, 18 hours on a 8mb apollo
3000.

as mentioned in the message that didn't drop into the bit bucket, the explorer is
faster than a lambda. the lambda however had a much better disk.

    Don Mitchell	  		Don@atc.bendix.com
    Bendix Aero. Tech. Ctr.		Don%atc.bendix.com@relay.cs.net
    9140 Old Annapolis Rd.		(301)964-4156
    Columbia, MD 21045

<dp>

From: Jeffrey Del Papa <dp@JASPER.PALLADIAN.COM>
Subject: more on workstation speeds
To: ai.works%osu-20@ohio-state.ARPA
Cc: johnson%ti-csl@csnet-relay.arpa

  LMI versus explorerI. according to RG (R.Greenblatt), the difference in speed
between the two processors is in rough proportion to the processor cycle time. the
lambda clocked it's processor at 200ns, the explorerI at 140ns. (the symbolics uses
either 100ns or 200 ns, it is hard to tell from the stuff they made public, but since
the instruction set and data path's are so different, comparing raw clock speed isn't
enough)

  The explorer II runs it's clock at 25mhz (40ns) in the currently shipping version,
with some of the chips doing fine at 30mhz (33ns) (and this is from the third run
(first non prototype) of chips) I was suprised to find out that the chip is CMOS with
a silicon substrate (CMOS on saphire is faster, but a bitch to produce) with 1-2
um features (read: newish, but not particularly exotic production requirements)

  Contrary to Don Mitchels report, the explorerII does not have a new disk
controller, or disks. (they were going to upgrade to the 190mb maxxtors, but the
larger capicity drives are too shock sensitive, and the QA people won't let them use
them until they pass muster) 6-8 months ago they did start selling a SMD controller
and drive, (tho we didn't get one for our trial) which would make things even faster.
(maxxtor 140 drives and the scsi interface are blindingly slow, compare a 3620 to a
3650 on a problem big enough to page, the only difference between the two is the disk
drives (and box))

  The biggest negative of the explorer(s) (of which the lambda is also guilty) is the
small address space, (yes we have bumped into it) followed by the fact that single
floats dont fit in one word. (both of our products do a huge amount of floating point
calculation) 

  The Ivory chip isn't supposed to be ready until late next year, and I doubt it can
be that much faster. (even 68k's are only running at 35-37 mhz (lab speed) I haven't
heard of production chips faster than 30mhz))

  The explorerII list price is currently higher than the base 3620 system, (36k vs
45k) of course when you adjust memory sizes (the explorerII comes with 16mb) the
prices are not all that different, and TI is a lot more aggressive about discounts.
(basicly a TI salesperson has a pre authorized discount schedule, with 'bolics all
exceptions to list had to be appoved from above)

  I was initially somewhat embarassed to be coming to TI's defense, given all my
friends at symbolics, the 3640 on my desk, all the cursing I did when it was first
time to move our software onto their machine, (pre-release H of the operating
system), and the general sluggishness of a 4mb explorerI; but they have done a very
good job on both the new hardware, and the new release of the software, and while I
have complimented some of them privately, some public recognition is in order.

<dp>



-------- End Forwarded Message


further comments: I still think the explorer monitor is truly ugly (E.T. go home) and
I would like to get whoever spec'd the conectors for the back of the disk and tape
bricks alone in an alley sometime, (In most of the maint contracts that TI writes,
those connectors are specifically excluded) possibly along with the person who
decided to use fiber to the monitor. (it is not available in teflon coated, so you
can't run it thru the ceiling of a modern bldg. it also breaks if you give it a dirty
look, drop a computer floor tile on it, etc. you can't pull it thru conduit as the
connector is permanently attached.)

<dp>
<dp>