PowerPC in your future?
David G. Hough on validgh
dgh
Fri Oct 8 22:24:02 PDT 1993
Craig Hartley sent me the following. One line of rebuttal is that none of
the PowerPC operating systems are going to be of any overwhelming interest
(almost but not quite MacOS, Pink - IBM is having doubts, NT which is
almost but not quite Windows, AIX - little interest so far, Solaris 2)
and so the volume arguments won't hold water.
From my point of view it complicates life to have to plan software that
works on conventional single/double RISCS, extended-based x86's, and
fused-multiply-add PowerPCs.
Another point of view is that all RISCS will be equivalent, with DOS,
Windows 3.1, and MacOS emulations of various sorts that permit most existing
mass market PC programs to run on the RISC PC's and workstations.
It is clear that there isn't enough value added to the world's overall
welfare to justify continued aggressive funding for development
of all of x86, PowerPC, SPARC, PA-RISC, MIPS, Alpha... on an indefinite basis.
===============================================================================
Ten Reasons Why PowerPC Will Succeed
1) Major Manufacturers' Commitments Already Guarantee Its Market Success
It will have by far the largest installed base of any RISC processor in
history. Indications are that it will pass 1M units shipped in 1994,
making it the first RISC processor to pass 1M in a year (SPARC
reportedly has around 400K/yr world-wide). As if being the standard
processor for both IBM and Apple weren't enough, it has also been
adopted as the standard embedded processor for all Ford autos (maybe
two per car). It will also find its way into TV cable boxes and sets
through alliances between Apple and Motorola and various media
companies. Computer ResellerNews on August 23, 1993 quoted InfoCorp as
forcasting PowerPC market penetration as follows:
YEAR SYSTEMS SHIPPED % OF TOTAL RISC PCs
---- --------------- -------------------
1993 497 <1%
1994 425.9K 42%
1995 2.39M 79%
1996 4.65M 83%
Note that this is conservative and reflects only "PC" sales, not cars,
TVs, and presumably not workstations and servers, etc. These numbers
are conservative relative to what Apple plans for 1994 (reportedly 1M
to 1.5M PowerPC units shipped).
Some would argue that embedded versions of PowerPC are unrelated to the
CPU version. My response is that the potential market for embedded
control is enormous. In-Stat Inc. (Scottsdale, AZ) forcasts the demand
for embedded solutions this year at 114.7 M - twice the size of the PC
market. By 1997, the market will grow to 203.2 M, or 2.5 times the
size of the PC market then (data from EE Times, 10/4/93, p 8). It
would seem that having the volume and intrinsic code compatibility
would leverage and strengthen the PPC market position when compared
with competitors with a small market share and no embedded versions.
Note that SPARClite by Fujuitsu is an embedded processor going after
this market in competition with PPC and Intel has a new "extended 386
series" with competition from AMD's Elan processor.
Another comment I've heard said that 1M installed SPARC systems have
more relevance than analysts' projections. My response to that is that
it took 5 years for SPARC to get that installed base, PowerPC will
probably achieve 1M+ installed base in less than 1 year. I have seen
no basis for doubting those projections. In fact I think they are
*very* conservative.
There is a question about whether Apple PowerPC 62/72 SPECmark machines
are really any high-end competion, because they don't make
"workstations" like Sun, for example. I think that the term
"workstation" will become practically meaningless soon (if it isn't
already). What is a "workstation"? Unix?...fast
CPU?...ethernet?...hi res graphics?...lots of memory?...big hard
drives?...multiprocessors? The new PowerPC systems will have all of
these things plus, potentially, more usable operating systems and a
huge number of easy-to-use applications (because of ports to a volume
architecture). As the saying goes, "If it walks like a duck, and
sounds like a duck, and looks like a duck...then maybe..." And what
about IBM's new PowerPC-based "workstations"?
2) It Can Beat the Competition in Price
For example, if you compare the available revenue per wafer, the
PowerPC has a continuing price advantage over Pentium, according to a
table on page 116 of the 9/20/93 MacWeek:
POWER POWER DIES PER
DIE AREA SUPPLY DISSIPATION 8" WAFER $/WAFER
--------- ------ ----------- -------- -------
Pentium I
0.416 Sq in 5.0 V 14.0 Watts 114 $34,200
Pentium II
0.235 3.3 V 6.5 Watts 203 $60,900
PPC 601
0.185 3.6 V 8.5 Watts 258 $77,400
PPC 604
0.128 3.3 V 6.0 Watts 373 $111,900
DEC and MIPS are trying hard to challenge the PowerPC but it is very
unlikely they can hope to achieve the volume that will drive prices
down as far as PowerPC can go, assuming all manufacturers are getting
equal yields. That may be a big assumption: Pentium has apparently
been having problems; PowerPC is technologically more advanced (4 layer
metal process) and so may have trouble ramping up yields.
For comparison, a current 486DX2 at 66 MHz goes for around $500,
PowerPC 601/66 will be around $280 in volume. Think of what a PC clone
costs now and you get the picture.
3) It Is More "Open" Than Any Processor on the Market
The PowerOpen Application Binary Interface (ABI) for PowerPC
incorporates the Mac API (toolbox) and indications are that the chip
has the ability to run the Macintosh OS (version 7.5) in emulation at
speeds roughly equivalent to the current high-end Macs (25 MHz 68040).
IBM signed up for Wabi for it's RISC 6000 product line (which is
transitioning to pure PowerPC). In addition, IBM is rumored to have a
"clean room" project to put DOS/Win code on PowerPC to speed x86
emulation (see original article below). The chip is so fast that even
in emulation it may well produce acceptable performance, and with the
installed base growing rapidly, most software will port to native
PowerPC very soon or lose market share to those who do. Remember Apple
& IBM are behind this and offer a guaranteed market. IBM
single-handedly made the x86 a success by merely adopting it, despite
the fact that the 68xxx was superior in many respects (remember those
pre-SPARC Suns?). I define "open" not as a published standard, but
rather as really being able to use as many off-the-shelf hardware and
software elements as possible (e.g., any brand of film in any camera on
any tripod). PowerPC will run Solaris, AIX, NT, Macintosh, MS-DOS, and
whatever IBM has as its office environment. Now, that's "open"! If
software emulation and hardware features give compatibility with the
installed base, and recompiled performance is very high, the pain of
the changeover could be acceptable.
4) PowerPC Will Incorporate PCI in its Architecture
Future versions of the PowerPC will integrate PCI bus support into the
chip design. Apple and IBM will standardize on PCI. This is no big
deal in itself, because Alpha already integrates PCI and MIPS probably
soon will, but PowerPC at least meets the competition.
5) IBM is Aggressively Marketing PowerPC to Asian Clone Manufacturers
EE Times reported last week that IBM is close to making a deal to form
a consortium including 85% of the motherboard manufacturers in Taiwan.
Those manufacturers are distressed at their inability to acquire
Pentium chips from Intel. IBM is promising them plenty of supply. If
these manufacturers can ramp up production (IBM will be presenting a
basic sample design at COMDEX), you may be able to buy a PowerPC-based
clone at Frys for under $2K in the not too distant future. That system
would run at least 62 SPECint92 and 72 SPECfp92, possibly 2 to 3 times
more! IBM/Apple is trying to capture a significant share of the RISC/NT
market from its rivals at DEC and MIPS. Capturing market share, not
only for NT but against x86 is important for what I think is really
underneath the strategy...see #6 below.
6) PowerPC will be the Platform of Choice for the First Object
Oriented OS to Hit the Market, Taligent's Pink
Indications are that Taligent's Pink, the product of another IBM/Apple
alliance, will beat Cairo to market. Taligent will reportedly be
selectively showing Pink to folks at next month's COMDEX. Microsoft is
showing Chicago (Windows 4.0, one generation preceeding Cairo), which
won't ship until late 1994. It would appear, therefore, that Cairo
most likely will follow Pink. If IBM is successful in getting PowerPC
clones rolling out of Asia, combined with Apple and IBM production,
PowerPC will have a huge installed base ready for the new OS. All
Taligent has to do is offer it for $99 "competitive upgrades" and you
have an overnight success (let's see...by mid 1995 3M cpu's /2 (half
the people might not hear about the deal) x $99.00 = $148.5M almost
overnight). If Pink employs some of Apple's slick technology
(QuickTime, speaker independent speach recognition, etc.) it could be
almost irresistible, even at a higher price. Taligent could become the
Microsoft of the late 90's off the PowerPC springboard.
7) PowerPC Systems may be More Usable than those of the Competition
I am intrigued by the notion of not price vs performance, but rather
usability vs performance. I started my PowerPC file many months ago
precisely because of my interest in this assessment. (I am a Human
Factors Engineer.) Unfortunately, I don't have hard data to back up my
assertions in the figure, much as I'd like to as an engineer. I throw
this out as food for thought. I know there are lots of ways to define
usability, and there are lots of excuses for not having it. I accept
the different definitions, but as an HFE, I can't accept any excuse for
not having it (e.g., "...a complex, highly functional OS can't have a
simple, friendly interface"). Usability can be achieved if there is a
commitment do do so. Perhaps someone might have a better idea of
relative ease-of-use for different operating systems. If so, send
feedback directly to me and I'll update my figure. The idea behind
this figure is that the design target is the upper right corner, i.e.
high performance that is easy to use. All the performance in the world
will only get you a limited market share if it takes a rocket scientist
to routinely get anything useful out (e.g., Unix). On the other hand,
a truly easy-to-use system will create new markets and opportunities
(e.g., Macintosh and dektop publishing). This crude figure will give
you the idea (and will serve as a classic example of the tool
limitations of the current state-of-the-art system I'm using to write
this; new OSs and standards will likely offer graphics, sound, and
video in their e-mail environments).
_________________________
10 | | |
| High P | Design |
| Low U | Target |
P | | |
O |___________|____________|
W | | |
E | Low P | Low P |
R | Low U | High U |
| | |
|___________|____________|
0 10
USABILITY
With (hypothetical, projected to late 1994) data filled in:
_________________________
10 | | |
| | 4 |
|1 2 |
P | | 3 |
O |___________|____________|
W | | |
E | | |
R |59 | 6 |
| | 7 |
|___________|________8___|
0 10
USABILITY
1 = SPARC-based UNIX system
2 = SPARC-based UNIX with Motif GUI (COSE)
3 = Pentium with WindowNT or Chicago
4 = 80 MHz PowerPC with Taligent Pink (late 94?)
5 = 486 standard Unix system
6 = 486 with Windows 3
7 = 68040 Mac with System 7.x or PPC in Mac emulation mode
8 = 68030 Mac with System 7.x
9 = 486 with DOS
Off the scale very high and to the right (30/9) = PowerPC 620 with Pink
The bottom line is that PowerPC's user-friendly parentage (Apple) will
most likely make its native OS (Pink) a formidable contender, if not
the leader, in usability comparisons. I believe that the marketplace
of the mid 1990's and beyond will no longer tolerate unfriendly
software and hardware. Too many good examples will be around, hence
the operational costs of unusable systems will become glaringly
apparent. Band-Aids like COSE won't survive against new systems
designed from the ground up. Pink must eventually be ported to Intel
to truly succeed, but starting out on a PowerPC might help the OS take
hold. Also, PowerPC's superior floating point performance relative to
Pentium may make Pink consistently function better on PowerPC. The
trend is toward all OSs running on all microprocessors. Nevertheless,
my guess is that PPC/Pink will consistently appear high and to the
right.
8) History Is on PowerPC's Side
Historically, I believe I read somewhere that if a new system can offer
50% more performance at something like 75% of the price, it succeeds in
the marketplace. Using the current high-end 486 architecture as the
current baseline, PowerPC offers 2 to 30 times (in some demos) the
performance at less than 60% of the price.
9) NT Doesn't Matter Anyway
NT has a relatively small market with lots of competitors; the real
market is Chicago in the near term and the Object-Oriented OS of the
future (Pink & Cairo) in the long term. NT takes too much memory, like
Unix, to fit on most existing desktops, hence its potential market is
nothing like MS/DOS. I heard one expert (Micheal Slater of
Microprocessor Reports) say the total market for NT may only be 2 M,
which is a lot by Unix standards but small by PC standards. The NT pie
will be split among a number of competitors: Pentium, Alpha, MIPS,
PowerPC, 486, and someday maybe SPARC. The cost of memory will drop by
the time the Object-Oriented systems arrive, and those OSs may also have
built-in efficiencies and features that justify upgrades.
10) PPC Can Match or Beat Its Competition in Performance and Technology
To make this point, I offer my current processor file. A caveat is
that in reading this list, one must take into account the level of
integration of each chipset; some are almost complete and some need
assorted cache and memory controllers and I/O bus interfaces. In this
sense SPARC, for example, may be more competitive than it appears based
on SPECmarks alone. Also, one must realize that any new microprocessor
generation leapfrogs over the one before. What is important for the
PowerPC is that analysts like Michael Slater of Microprocessor Reports
believe that Motorola is *absolutely* committed to PowerPC for the long
haul. If the first generation PPC doesn't win they will keep on trying
and trying until they succeed in creating the x86 architecture of the
2000s. All data is from published sources like EE Times and product
literature. I put this together by reading during my lunch breaks over
the past 6 months and updating an e-mail (to myself) with any new
numbers as I found them. This is a week or two old, and I know a few
things need updated. I have also included some relatively outdated
processors (e.g., 386 and 68030) as a basis for comparison. I'd
appreciate any corrections being sent directly to me. I'll publish an
updated table if there is interest.
MICROPROCESSOR POWER AND COST
--------------------------------------------------------------------------
Microprocessor Clock Speed SPECmark Rating Cost
--------------------------------------------------------------------------
Alpha (DEC 21064) 150MHz 74.3 SPECint92 $853
30 Watts 126.1 SPECfp92
little-endian or
endian neutral
8k cache; 431 pins
Alpha (DEC 133) ? 65.3 SPECint92 ?
112.2 SPECfp92
Alpha (DEC 21064) 200MHz 106.5 SPECint92 $1,231 (QTY)
16K cache on chip 200.4 SPECfp92
4Mb cache off chip
Alpha AXP (DEC 21066) 166MHz 70 SPECint92 $385 (5K)
built-in MMU & PCI 105 SPECfp92
below 10 W ?
not multi-processor;
graphics controller;
287 pins; 8k cache;
integration eliminates
need for most system
logic parts
Alpha (DEC 21068) 133MHz? 56 SPECint92 ?
8.5 W 84 SPECfp92
--------------------------------------------------------------------------
PowerPC 601 66MHz 62 SPECint92 $374
(Motorola, IBM, Apple) 72 SPECfp92 $280 (QTY)
2.8M transistors
4-metal, BiCMOS
0.65 microns
die size: 11mm x 11mm
32K cache on chip
8.5 Watts, bi-endian,
32-bit address, 64-bit
data, 52-bit virtual
addressing
die area 0.185
3.6 Volts
demo at Apple Devl. 80MHz 75 SPECint92 ?
Conference (5/93) 87 SPECfp92
IBM plans to ship
systems at this speed
Q1/94
PowerPC 601 50MHz 47 SPECint92 <$280
6.5 Watts 54 SPECfp92
PowerPC 603 66MHz 62 SPECint92 (est) <$200
<4 Watts 72 SPECfp92 (est)
(by 12/93)
for portables
PowerPC 604 ? 160-180 SPECint92 (est) ?
(by 6/94) 210-240 SPECfp92 (est)
for desktop wkstations
performance estimates
based on press reports
of 3x 601
3.3 Volts, 6 Watts
die area 0.128 sq. in.
PowerPC 620 ? 250 SPECint92 (est) ?
full 64-bit version 330 SPECfp92 (est)
of PowerPC architecture
(by 9/94)
(420/560 SPEC possible)
PowerPC 400 (IBM) ? ? ?
32-bit embedded mp;
aimed at low-end PDAs
description due 12/93
design by IBM in NC
-------------------------------------------------------------------------
Power 6264 (IBM) 62.5MHz 59.2 SPECint92 ?
32K cache on chip 124.8 SPECfp92
64K cache off chip
RS4132 (IBM) 41.7MHz 34.6 SPECint92 ?
65.0 SPECfp92
------------------------------------------------------------------------
Pentium (Intel) 66MHz 64.5 SPECint92 $950 in vol
3.2M transistors 56.9 SPECfp92 $965/1000
0.80 microns
die size 16.8mm x 17.6mm
die area 0.416 sq. in.
16K cache on chip
116 mips
256K cache off chip 60MHz 58.3 SPECint92 $900 in vol
13-16 Watts 52.2 SPECfp92 $878/1000
little-endian
Pentium II (Intel) 66MHz 64 SPECint92 (est)
die area 0.235 sq. in. 57 SPECfp92 (est)
3.3 Volts, 6.5 Watts
0.60 microns
P54C-133 (Intel) 133MHZ 120 SPECint92 (est)
110 SPECfp92 (est)
P6-125 (Intel) 125 MHz 160 SPECint92 (est)
150 SPECfp92 (est)
------------------------------------------------------------------------
microSPARC (TI) 25MHz 13.8 SPECint92 ?
SPARCstation 1, LC) 14.0 SPECfp92
all current SPARC are
big-endian
40MHz 19.6 SPECint92 ?
21.2 SPECfp92
SPARCclassic 50MHz 26.4 SPECint92 ?
21.0 SPECfp92
microSPARC (TI) 50MHz 26.4 SPECint92 $179 (QTY)
(SPARCstation 2,LX) 21.0 SPECfp92
6K cache on chip, none
off, includes mem
controller & SBus
interface
aka "Tsunami"
microSPARC II (Fujitsu) 75-100 MHz ? ?
24K cache on chip,
none off, icludes
memory controller
and SBus interface
Power up (Weitek) 80MHz 32.2 SPECint92 $1500 (one)
clock doubling chip 31.1 SPECfp92
for SPARCstation2 &
IPX; 24kbyte cache;
external 40MHz clock;
pin compatible w SS2
and IPX
SuperSPARC LC (TI) 40MHz 45.2 SPECint92 $299 (10K)
(SPARCstation 10) w/o mxcc 54.0 SPECfp92
36K cache on chip
1Mb cache off chip 40MHz 53.2 SPECint92 ?
aka "Viking" with 1MB 63.4 SPECfp92
0.8 micron BiCMOS
SuperSPARC+ (TI) 50MHz 65.2 SPECint92 $549 (10K)
(SPARC10 Model 51) with 1MB 83.0 SPECfp92
SuperSPARK+ (TI) 60MHz 77-80 SPECint92 $849 (10K)
with 1MB 95-100 SPECfp92
SuperSPARC II (TI) ? >100 SPECint92 ?
1994, dual-launch FPU to 200 SPECfp92
HyperSPARC (FMI) 55MHz ? ?
16 KB on chip cache
now modules only
256 KB off chip 66MHz ? ?
combined;
dual MBus modules
Thunder (Hyundai) ? ? ?
metaflow design
microSPARC II (FMI) 50-60 Specint92 ?
3.3 v, low power 92 SPECfp92
for laptops, embedded
apps; announced 5/93
ultraSPARC I & II ? ? ?
64-bit, CMOS .5 micron,
four-scalar, four-level
metal, dual launch FPUs,
uni- and multi-processor
support
bi-endian, 1995
ultraSPARC III 700-1000 SPECint92 ?
64-bit, BiCMOS .4 micron
bi-endian, 1997
------------------------------------------------------------------------
PA7100 (HP) 99MHz 80 SPECint92 ?
no cache on chip 150.6 SPECfp92
512K cache off chip
bi-endian
PA-RISC 1.1 (HP) 66MHz 51.1 SPECint92 ?
75.0 SPECfp92
PA-RISC 1.1 (HP) 50MHz 32.8 SPECint92 ?
47.6 SPECfp92
------------------------------------------------------------------------
R4400 (MIPS, SGI) 75MHz/150MHz 94.2 SPECint92 ?
32K cache on chip 105.2 SPECfp92
4Mb cache off chip
bi-endian
R4000PC (MIPS, SGI) 50MHz 34 SPECint92
standard in Indy 35 SPECfp92
100 MHz internal clock
R4000SC (MIPS, SGI) 50MHz 58.3 SPECint92
option in Indy 61.5 SPECfp92
100 MHz internal clock
R4000 clones
2-3 Watts, NT available ? ? ?
-----------------------------------------------------------------------
80486DX3 (Intel) 100MHz 48.3 SPECint92 $500
33/100MHz (clock 24 SPECfp92
trippled); 12/93
80486DX (Intel) 66MHz 32.2 SPECint92 $500 (est.)
little-endian 16 SPECfp92
54 mips
33Mhz 16.1 SPECint92 $264 (ave.)
8.5 SPECfp92
80486SX (Intel) 33MHz 11 SPECmarks $64 (Ave.)
80386 (Intel) 33MHz 3 SPECmarks ?
-------------------------------------------------------------------------
Am486DX (AMD) 33-40MHz clone 486 $306
little-endian
Am486DX2-50 (AMD) 50Mz clone 486 $417
clock doubled version
--------------------------------------------------------------------------
68060 (Motorola) ? 60 SPECmarks (est) ?
0.5 micron, 2.8M (Pentium class)
transistors, 5 watts,
3 volts
68040 25MHz 18 SPECmarks ?
big-endian 33MHz 23.7 SPECmarks
40MHz 28.8 SPECmarks (41.8 Mips)
68030 25MHz 4 SPECmarks ?
big-endian
-------------------------------------------------------------------------
Clipper (Integraph) ? ? ?
(no data available)
Please send comments and corrections to: 70214.1036acompuserve.com
Craig Hartley
More information about the Numeric-interest
mailing list