[NTLUG:Discuss] Alphaserver 2100 -- Digital +3y, AMD v. IBM (Intel irrelevant), lesser chips

[Justin M. Forbes]

On Fri, May 21, 2004 at 08:43:53PM -0400, Bryan J. Smith wrote:
> 
> - MIPS:  Still going strong
> 
In the embedded market they are not what they once were, but not bad
either.  In the server market they are being phased out faster than I would
have expected.

> 
> - IBM:  Knocked out, now back with a vengence
> 
> IBM had a _total_failure_ with their original PowerPC 620 design back in
> '97-98.  So IBM created the 64-bit Power4 without any 32-bit PowerPC
> compatibility.  It wasn't until about 2 years ago when they announced
> the PowerPC 970 that they re-established 32-bit PowerPC compatibility. 
> It is what the Apple G5 uses.  IBM, of course, has the leading fabs of
> 
This is all completely incorrect.  Power 4 has always had 32bit
instructions, and AIX 5.x offers a 32bit kernel which runs on all current
generation P-Series hardware should you be inclined. (or forced to run
oracle 8i still)  One reason if confusion, most software vendors offer
32bit versions for AIX 4.3.x and 64bit versions.  They also will offer a
64bit version for AIX 5.x. The reason you see this, and not 32bit versions
for AIX 5.x is the 64bit code required a few changes between 4.3.x and 5L,
so 64bit apps are not compatible between the two versions.  32bit apps are
compatible, so the 4.3.x version is supported on 5.
PowerPC 970 differs from Power 4 in a couple of ways.  Power 4 is a dual core
design, PowerPC 970 is single core, and has less cache.  This makes it
considerably higher yield (read cheaper to produce).  Second, and quite
important for Mac fans, the PowerPC 970 added the AltiVec instructions from
the previous generations G4 chips.  These are not 32bit compatibility
instructions, AltiVec is actually a couple of 128bit registers, a vector
execution unit, and a few new instructions.  Think MMX but much better.
Also some of the RAS features are missing, but that is a minor point.  and
the bus is different.
Power 5 adds SMT and a few other things, definite improvement over the
already excellent Power 4+.

> - Sun:  Phasing itself out of the game
> 
> Sun's UltraSPARC (SPARC v9) finally hit 64-bit in '96 I believe.  It's
> actually a good design.  Sun uses TI as their foundary.  They are still
> stuck around 0.15um feature sizes, so the UltraSPARC III cannot keep
> up.  It looks like Sun is getting out of the IC design business soon. 
> The UltraSPARC IV is a multi-die chip based on the US3, and they've laid
> off about 2,200 people in their semiconductor engineering division. 
> They're probably going to phase out to x86-64 over the next 5 years.
> 
Sun has many more issues than just the 0.15um fab.  Ultrasparc III would
have been great, if it came out 2 years earlier.  US IV is simply dual core
US III, and the US V team got the axe just after tape out I believe.  There
are a couple of options for the future of Sun.  They are selling Opteron
based servers now, and supposedly are working with Serverworks on an 8 way
chipset, for general use, and a 16-32 way chipset to which Sun would have
exclusive rights for a certain period of time (unknown how long).  Opteron
has its own interconnects up to an 8 way, but to go beyond that, you need a
good high speed interconnect, and I think the Sunfire technoligy is being
looked at there.
The other option is partnering with Fujitsu whose current Sparc64 V
competes quite well with Power 4, much faster than current Ultrasparc. I
have heard a few things about the politics here, and I will keep them to
myself, but do not leave it out of the realm of possibility. Fujitsu's
chips are solid, if not commodity.  Even this would not be a saving grace
for Sun though, as I do not see the Sparc64 being able to compete on price
for very long.

> - Future:  AMD v. IBM
> 
> The future is going to be built on two major platforms.
>  - AMD x86-64 (aka Intel E64T)
>  - IBM Power4/5
> 
I would agree here.

> AMD has clearly taken over Intel's spot.  x86-64 is more than just an
> ISA, but AMD's HyperTransport and on-board I/O MMU design is years ahead
> of what Intel has planned for it's aging Pentium series.  The
> "extensions" Intel offers in the Prescott core are just compatibility,
> and not very competitive -- especially from an I/O-throughput
> standpoint.  

This is highly dependant on application, and hyperthreading (SMT) plus high
clockspeeds will offer performance in some areas that Opteron will not.
However Scalability, I/O throughput, NX (security) and many applications which
can benefit from AMDs excellent prefetch still leave AMD at the top for
now.  In the future, Intel is dropping the netburst architecture, and
moving to dual core.  AMD will have dual core Opterons next year, and they
will supposedly work in any Socket 940 motherboard you buy now. I would say
AMD has the lead right now in technology, but they will have to fight to
keep it.

> IA-64 Itanium III still doesn't offer the
> bang-for-the-buck, and no IA-64 probably ever will, even at native
> execution sadly enough.
>
IA-64 is supposedly going to be priced in line with  Xeon and share system
boards even by 2006 they say.  Bang for the buck will be there at that
price point for certain applications. The problem with IA64 is it can
perform quite well, but it has to be optimized for parallel execution at
compile time.  gcc does a horrible job, the Intel compiler does a better
job, but hand optimization is required for IA64 to be competitive.  For
general purpose, I do not see it happening, but for HPC, it is a
possibility.

> The other challenger is IBM Power5.  The current Power4 is an excellent
> example of what to expect out of Power5.  IBM will release a 32-bit
> PowerPC compatible version of its Power5, just like they did with the
> Power4 in the PowerPC 970 (aka Apple G5).  Where Power kicks x86-64's
> but is in efficiency.  It's half as complex as Athlon64/Opteron or
> P4/Xeon-E64T.  This is especially ideal for both "Blade" servers as well
> as portables.
>
I am not sure where you get this efficiency bit, Power is actually less
efficient than Intel in power consumption per transistor, even at 90nm.
PowerPC 970FX should run close to 55W, which is not bad for a server chip,
and much lower than Intel's number (and about half of intel's transistor
count).  Opterons do offer 55W and 35W variations though and that is
nothing to sneeze at. Interesting as well that you would use the term
complexity here, as Power is notorius for its complexity.  Don't get me
wrong, I think the Power architecture is excellent, and will be competitive
for a long time, but let's not stretch the truth.

> P4 mobile processors have to "slow down" to under 1GHz when on battery,
> to keep consumption under 20W -- they normally use over 50-70W at full
> 2+GHz speeds.  Even the now 18 month old PowerPC 970 at 1.2-1.4GHz
> (which is equivalent to more than a 2GHz P4) only used 19W.  IBM held
> off on releasing the part, and concentrated on 1.8-2.2GHz (using 35-45W)
> parts for Apple G5 systems and its own workstations instead, waiting on
> the 0.09 versions to come out with faster speeds (1.6GHz+) and lower
> power (close to 15W).  But, again, yield problems are holding the
> release up.
> 
Funny how you chose to use full die (core + cache, etc) power numbers here
for Intel, and only core numbers for Power.  In reality, the chips (core +
cache, etc) are drastically close in P4 vs. 970 vs. AMD64.  It is prescott
which comes with a massive transistor count increase, and thus power
requirement.  The mobile line (which in dual core may well be the succesor
to the failaing netburst P4/Xeon line) is not so bad at 90nm.

> IBM is opening up PowerPC licensure quite liberally.  So it may be
> difficult for AMD and Intel to hold onto marketshare by the end of the
> decade.  Especially if Linux catches on -- which is what IBM is hoping
> for.  Intel's IA-64 has become irrelevant, as HyperTransport is
> available in an optical flavor that scales far better for Opteron than
> Intel can do with IA-64.
> 
In the embedded market, you are correct, for the commodity desktop and
server levels, IBM has had issues getting anyone besides Apple to use PPC
970.  There is finally a reference platform system board available in case
motherboard makers want to make 970 dual workstation boards, but I have not
even heard a volume product announcement.  No question 970 is a nice CPU,
but I think it faces a tougher battle for mindshare then either Intel or
AMD.

Justin