🚀 go-pugleaf

In article <1992Nov12.091854.22914@walter.cray.com> cmg@magnet.cray.com writes:
>
>We must be careful when comparing new architectures with old
>architectures because of the effect of software.  [...]
>		[....]
>The improvement of microprocessor speed of the past decade has been
>tremendous.  How much of this improvement is attributable to software?

Without bothering to hunt down the numbers, it is clear that for the
LINPACK 100x100 and 1000x1000 benchmarks, the effect of software
improvements has been tremendous.   Because of this, I no longer
consider any of the LINPACK numbers to be useful for characterizing
system performance (except for dense linear algebra).

The two cases have been improved by different sets of software
enhancements: 

(1) In the 100x100 case, inlining has been the real key, though there
is still some room for improvement in the compile-time evaluation of
the excessive logical predicates in the BLAS routines.  On the hardware
side, larger caches have helped quite a bit.

(2) In the 1000x1000, the improvement has been mostly due to the
industry's increasing knowledge about block-mode algorithms, although
the use of pipelined FPU's has been crucial for these algorithmic
improvements to be effective.


Just as in the case of the SPEC89 Matrix300 benchmark (which was made
useless by the combination of these two sets of optimizations), this
set of enhancements has made the LINPACK test cases rather poor
predictors of scientific workload performance.  For most large-scale
scientific applications, the real bottleneck continues to be
sustainable memory bandwidth --- often requiring non-unit or irregular
strides.

The combination of unbalanced architectures (fast and/or pipelined
FPU's combined with slow memory systems) with an algorithm that
benefits far more than most from blocking results in systems that
cannot deliver what these LINPACK numbers appear to promise.

----------------------------------------------------------------------

It is interesting to note that the top-of-the-line DEC Alpha system
appears to be about even with the Cray C90 on $/MFLOP for the LINPACK
1000x1000 case.   Since this is a best case for the Alpha, one must
conclude that the C90 is more cost-effective for more memory-intensive
vectorized algorithms (i.e. most all of them).

	Alpha AXP 10000 : $ 300k/111 MFLOPS = 2.7 k$/MFLOPS
	Cray C90 - 1 cpu: $2500k/871 MFLOPS = 2.9 k$/MFLOPS

The Cray price is a guestimate.  It is probably not too far off
for a uniprocessor system.  Multiprocessor systems are likely
noticeably cheaper per cpu.
----------------------------------------------------------------------

It is also interesting to note that the DEC AXP 3000/500 at 150 MHz
has almost identical LINPACK 100x100 and LINPACK 1000x1000 numbers
with the IBM RS/6000-970 at 50 MHz.   Prices are similar, with DEC
appearing to have a slight edge: $39k vs ~$55k.

----------------------------------------------------------------------

All in all, a disappointing set of announcements from the point of view
of this number-cruncher....
-- 
--
John D. McCalpin                        mccalpin@perelandra.cms.udel.edu
Assistant Professor                     mccalpin@brahms.udel.edu
College of Marine Studies, U. Del.      John.McCalpin@mvs.udel.edu