[Info-vax] BOINC for VMS

[John Wallace]

On Apr 3, 6:10 pm, David Froble <da... at tsoft-inc.com> wrote:
> John Wallace wrote:
> > Whether Kumar thinks Intel are stupid or not is largely irrelevant
> > anyway. The historical fact is that Intel in recent years is a one-
> > technology company, and that technology is x86. More specifically,
> > it's implementing ever more complex x86s by enhancing their chip
> > fabrication technologies. x86 hasn't got faster for years, you just
> > get more of them on one core. Outside the x86 sector, they haven't had
> > a significant commercial success for years, and even in the x86 sector
> > they have to resort to anti-competitive tactics to "motivate" the
> > system builders to stay with Intel rather than the alternatives.
>
> We're all entitled to opinions ....
>
> Your statement about x86 not getting faster needs to be examined.
>
> If you mean that the architecture hasn't been getting fundamental re-design, you may have
> a point.  But "not faster", would not be correct.  Every time they do a process shrink
> there is speed improvements.  Basically what you alluded to in "enhancing their chip
> fabrication technologies".
>
> Since the end of the Alpha, process shrinks have been one of the major reasons for speed
> improvements.  AMD, and then Intel (I believe) moved to on-chip memory controllers, but
> that was first done on Alpha EV7.
>
> The windoz system I use has an Athlon 64 San Deigo core.  I once wrote that it was the
> fastest single core CPU ever mfg, and would remain so.  Got a lot of ridicule for that,
> but, I still think that I'm right.  Why?  Because both AMD and Intel declared that all
> future CPUs would have multiple cores on a chip.  SO it was indeed the last and best of
> the single core CPUs.
>
> You may indeed purchase a "single core" CPU, but that's because additional cores on the
> chip are either defective, or have been disabled.  AMD's Calisto CPU is a good example.
> Basically a Deneb 4 core chip, with only two (2) of the cores active, to fill a marketing
> niche.  Was a great gamble for many people, purchasing a 2-core chip, and unlocking the
> additional cores.  A gamble, because sometimes the additional cores were good, and
> sometimes not.  I got one, could use 3 cores, but the 4th was broken.
>
> At that time the CPUs were hitting a "wall".  There is a finite speed for electricity, and
> without possibly super-conductors, you cannot get any faster.  Process shrinks provided
> for shorter distances, and therefore "faster" CPUs.  What else to do?  More cores on a
> chip, and perhaps methods for better inter-core communications.  That's what you sometimes
> get these days.
>
> Myself, (which probably doesn't mean much), I feel that the OoO concept in Alpha and Power
> provides a method to get faster processing out of a set of instructions, if, IF, you can
> coordinate it well enough.  How to do this?  Simple, one word, "money".  Takes lots of
> smart people trying different ideas, adopting those that work, and setting aside those
> that don't work, "today".  DEC didn't, or couldn't, continue to fund Alpha.  So far I
> haven't heard that IBM has taken that route.
>
> Process shrinks and tons of cache have made the itanic much faster than the Alphas.  That,
> and some small inter-core communication improvements are all Intel has been able to manage.

Wrt out of order execution, DEC's "Alpha vs IA64" white paper
(published 1999) has some interesting and not too technical
discussion. Still available at
http://www.cs.trinity.edu/~mlewis/CSCI3294-F01/Papers/alpha_ia64.pdf

"Your statement about x86 not getting faster needs to be examined."

Fair enough. Facts are good.

""not faster", would not be correct. "

You reckon? Got the facts to support that?

Process shrinks can allow you to go to higher clock speeds, so long as
you don't hit other constraints. The constraint that killed Pentium IV
was power dissipation (or, in Intel-speak, TDP). But does its modern
replacement actually go faster? And if it does, what makes it faster?
Internal redesign at the same clock speed? More cache ? Faster clock
speeds? As far as I can see, Intel's x86 clock speeds for the most
part of the last decade have maxed out at around 3GHz give or take a
little, regardless of process shrinks.

Sensible people already know that comparing raw performance based on
clock speed alone is A Bad Idea anyway. Ever tried comparing Pentium M
vs Pentium IV from around the same era, for example - depending on
benchmark, you can get Pentium IV performance with a Pentium M at
around a half of the clock speed, with the Pentium M also running at a
much lower wattage. But that looks a bit weird in the marketing
material - you want to sell Joe Public a 1GHz processor and you expect
them to believe it'll be as fast as a 2GHz one ?

Rather than comparing clock speeds, someone looking sensibly at
performance might want to find an industry standard benchmark, ideally
one that's representative of the relevant workload. For a now and then
comparison, find some results from a few years ago, and their modern
equivalent, and see if there really is a significant improvement in
Intel x86 processor performance in recent years.

Ideally, to do this someone might be looking for a benchmark that
doesn't depend on IO performance, so something like SPEC's CPU2006
might suit as a starting point. Even better would be the option to to
isolate the influence of extra cache over the years - smaller
geometries in x86 have permitted more cores and more [total] cache but
not necessarily more raw processor performance. CPU2006 and indeed
much common software probably benefits from more cache, so a small
benchmark like the more recently introduced Coremark might be worth a
look.

Another factor to consider when comparing performance over the years
is what changes compilers have made; in principle better compilers can
often make the same source run faster even on the same hardware. gcc
has certainly improved in many ways in the last ten years or so.

I don't have the numbers to hand and need to be elsewhere...

Have fun!