-march=xxx

Nicholas Leippe nick at byu.edu
Wed Jan 24 14:31:26 PST 2001


On Wednesday 24 January 2001 10:28 am, you wrote:
> > > I
> > > dunno, IMO P4
> > > and Atlon families are more like the 6 and half generation.
> >
> > Athlon and Duron are 7th Generation CPU's in the Intel style.  It was a
> > major marketting point during the introduction of the Athlon.
>
>     Yeah, it's a pretty hotly contested topic over exactly what defines a
> processor 'generation'  In general I've seen it to mean a combination of
> significant change in the arch features of the processor above and beyond
> cache size/speed/bandwidth and a jump in the relative performance.  Also,
> historicly, there has been a large increase in main memory bandwidth with
> each suceeding generation.  Granted, this is more to do with the platform
> then the processor, but i kind of tie them together.  For example the
> Serverworks chipsets double and quadruple the main memory bandwidth by
> interleaving accross multiple DIMMS in paralell.  Still, it is considered a
> 6th gen platform.

I fail to see how interleaving across multiple DIMMS in parallel would
double or quadruple memory bandwidth on a point-to-point bus such as
Intel's pentium GTL+.  It could certainly reduce memory latencies (and
thus improve overall memory throughput) but the memory bandwidth is still
limited for the most part by the bus speed and width at the cpu's pins.

Processor generations have been generally defined and fairly well accepted
to consist of the following:

1st generation: (i8086, MC68000)
--------------
it works.
no cache
single in-order execution
no MMU
no FPU
no protected mode

2nd generation: (i80286, MC68020)
--------------
protected mode

3rd generation: (386, 68030)
--------------
added MMU
added cache

4th generation: (486, 68040, AMD 5x86)
--------------
added FPU
pipelined
increased cache
internal clock multiplier (core faster than bus)

5th generation: (pentium/MMX, 68060, PowerPC, AMD k5, Cyrix had one too)
---------------
superscalar (two pipelines--multiple issue)
increased cache

6th generation: (ppro/PII/PIII/celeron, PowerPC G3/G4, AMD k6/K6-2/K6-3/Athlon, Cyrix 6x68, 686MX, MII)
--------------
out-of order execution
register renaming
increased cache

see http://www.emulators.com/pentium4.htm for much more detail

>
>     When the Athlon was introduced the performance increase was impressive
> over the fastest available Intel offering at the time.  The platform it
> runs on is making it suffer.  It is based off good technology, but the
> execution is flawed.

I really fail to see how the Athlon design is flawed in comparison to
Intel's.  In most respects it is at least equal, if not better.

> Slow memory performance is just sapping Athlon of
> it's power. The performance gap has narrowed recently.  If you go through
> various benchmarks it is interesting to note that the Athlon is real good
> at certain types of tasks but only a little better at others.

I really don't know why you think the Athlon's 'slow memory performance' is
'sapping' its power.  Why do you claim that it has slow memory performance 
in the first place?  Its bus protocol (Alpha EV6) is actually superior
to Intel's GTL+ in many respects--especially in memory bandwidth.

> I will be real interested to see AMD's answer to the P4.

AMD won't have to answer to the P4.  It, in reality, of all designs is one
of the poorest of the lot.  It is inherently flawed, and doesn't answer
neither any technical challenges in any new way (it actually regresses to
older, inferior designs in a few cases), nor does it answer current
market demands by any means.  Intel has stuck themselves between a rock
and a hard place technically, and will have to use marketing brute force
to get out of it at this point.

AMD's next iteration of the Athlon core--the Palomino and friends, will
easily crush the P4.  (heck, the current Athlons already do)  And the new 
cores don't add any great big sweeping architectural one-ups to do it.  
For the most part it'll be just a die shrink and larger cache.  The Athlon 
core is still in it's infancy of product line and growth potential, 
compared to Intel's p6 core which is on it's last leg--they can only hope 
to squeeze just a little more out of it with die shrinks and speed bumps.

>     Another thing is the ammount of performance increase through-out the
> 6th generation is just incredable.  Remember, it started with the lowly
> PPro 166Mhz and now at 1.13Ghz with the P3.  AFAIK, Intel is still planing
> a die shrink to .13 micron with the P3 core.  Maybe it'll make it up to
> 2Ghz, still a member of the 6th Gen.

Perfomance increase is not directly proportional to core clock speed!!!
The megahurtz race is futile and misleading.  Go read up on Hennessy and 
Patterson if you need details.  But, your are correct in that the current 
1GHz cpus give better performance over the 166 (actual first ppro was 150) 
ppro. (And the 1.13 GHz p3 doesn't exist in functional form--that exceeded 
the limits of the .18 process with that core and had to be recalled.

>     So, Athlon has some significant Arch. changes, increases performance in
> some, but not all, areas, and no significant change in memory bandwidth
> (OK, DDR will help, but P3's gonna have that to)  IMO the true 7th Gen will
> be 64-bit with Sledghammer and the final release of whatever
> itanium/McKinnly whatever gets released.

64-bit Sledgehammer may be called 7th generation, depending on what other
architectural changes they make besides a straight-forward register width
extension.

ia64, otoh, is an entirely different architecture, and doesn't really
pertain to any generation of current cpu families.  It's not CISC, and it's
not RISC.  It's a variant of VLIW, of which there have been/are only a few
other archs that did so as well (I believe the CRAY vector cpu was a vliw,
but not sure).  Intel is having a real struggle with this one, and it's
not looking like they're gonna have their head above water very soon with
it either.  And that's just aiming at the server market for now anyways.
They don't have much of a prayer at this point of trying to penetrate the
consumer market with it.

Nick

-sorry about kicking the dead horse, but it was still twitching :)

-- 
Nobody will ever need more than 640 kB RAM.
                 -- Bill Gates, 1983
Windows 98 requires 16 MB RAM.
                 -- Bill Gates, 1999
Nobody will ever need Windows 98.
                 -- logical conclusion

-- 
Unsubscribe: send email to lfs-apps-request at linuxfromscratch.org
and put unsubscribe in the subject header of the message




More information about the blfs-support mailing list