Grumble wrote:
Robert Myers wrote:

The _other_ company sells performance. We sell dreams. :-).


My, oh my. Are we in a trolling mood or what!

:-)


The implication being that selling dreams over performance is a bad
thing? I don't happen to think it is.

Intel's Chief Architect of the P4 era says that he warned Intel
management that the NetBurst architecture had bought them some breathing
room--a few years at most.

The players in the Intel architecture mini-drama seem to an outsider as
if they might be stubbornly goal-oriented, but they don't, any of them,
seem inept.

Intel sometimes seems to act as if it could rearrange virtually
anything, including physics, to suit its market objectives. Anybody
knows that, in a showdown between physics and marketing, physics wins.

Except if you're Intel, apparently. :-).

Over the long haul, if people really need performance, and Intel can't
deliver it, Intel is in trouble. On the short haul, it would be
difficult and dangerous for Intel to try to persuade people that it is
silly for them to focus on performance differences they will never
notice. In a year or two, Intel might be right back where it was with
P4, trying to convince them that a 2.4GHz CPU was faster than a 1.7Ghz
CPU in a way that justified an upgrade. Right now, though, Intel needs
to get its customers to think about something other than single-threaded
performance without engaging in tedious and dangerous explanations.

Intel has big architectural changes in mind: offloading significant
pieces of work, like network processing, from the main CPU. If Intel
were giving advice in "The Graduate," it would be whispering
"Multi-threading." Intel has the clout to make a major change in
programming style like that stick, but even Intel can't make it happen
overnight. In the short haul, Intel has to concede single-threaded
performance to AMD and to get its customers to think about something
else. It looks to me as if they understood exactly what they need to
do, and they are doing it.

Eventually, Intel will be back to selling performance, but the kind of
performance it will have to sell is going to require significant
customer education.

The consumer CPU business may repeat the mistake of HPC and try to force
a vector quantity (usable performance) into a nearly meaningless scalar
(linpack, top 500, etc.), but I can hope not. With any luck, Intel will
be able to move the focus off single-threaded performance and things
that look like stand-ins for single-threaded performance onto more
sophisticated measures of value.

Sophistication in measuring value is great conversation for hardware
groups on Usenet, but when you're trying to make a sale, you don't want
the party reaching for his or her credit card to be thinking about
complexity. In such a circumstance, dreams are the preferred commodity
to be offering for sale.

RM