Weak Signals and All That Kind of Stuff

[Mike O'Dell]

Walsh functions and their transforms have long been the 
technology of choice in Soviet-era signal processing systems
because they require nowhere near the computational capacity 
of the more famous Fourier orthogonal functions for implementing
DSP-like operations.

In fact, Walsh-Hadammard transforms can be implemented
in remarkably simple combinatoric logic.

I have a translated Russian book on communication system
design which spends quite a few pages on W-H orthogonal
functions and have seen another one which is focussed
almost entirely on "everything you'd ever need to
know about W-H transforms".

Back in the middle 1980s when I spent some time reviewing
(translated) technical literature from the "Eastern Bloc",
i was struck time and again by the difference in design
solutions they pursued vs what was au courant in "the West".
"Just because you have the technology to do something the
most impressive way doesn't mean you're doing it the *best* way."

Intel is suffering from this right now.
When binary compatability with the x86 instruction set
was an axiom because of the desktop hegemony of Windoze,
the manufacturing capability to throw at least a half-billion more 
transistors at a CPU design than anyone else in the world
let them stay ahead just on their frabrication technology.

Now that said OS hegemony is rapidly receding into history,
the instruction set of a processor is no longer the driver
it once was. Now it's Power and Cooling efficiency running
multiple cores that is the figure of merit. The ARM instruction
set is an order of magnitude less complex than the x86,
and approximately the same relationship still holds for 
the x86-64 and the ARM-64 ISAs. the difference in the
implementation, however is more like two orders of magnitude
for equal single-thread performance. That means a lot more 
cores go on the same chip with a lot fewer transistors.
That means less power for the same useful work and it means
much cheaper chips since a smaller part means more per wafer,
and larger geometry means better yeild and dramatically 
less processing fixed costs (multi-mode diffraction masks,
deep-UV steppers with diffraction optics, etc, etc.).
But most of all, the fact that ARM licenses their technology
far and wide, it means *competition* the likes of which
the computer industry has not seen in a quarter of a century.

It's not quite as bad as what happened to slide rules
literally overnight when the HP-35 appeared, but given
the exponentially larger dollars at stake, it's probably
worse in terms of several metrics one could chose.

Between what Intel has already bought and what they are
known to be shopping for, it's clear they are somewhere
between "terrified" and "scared shitless". The fact
that Microsoft just missed their numbers by the largest
margin in over a decade - with Windoze spouting arterial
blood loss - cannot be good news at Intel.

The Irony of Ironies is that Intel *had* the best ARM
design and design team in the industry when they got
the DEC silicon design team and the StrongARM/Alpha group.
They promptly killed the Alpha and stripped the new
StrongARM design of its floating-point and vector units
because it would have *crushed* the 386. After starving
the group into oblivion, the cremains of the intellectual
property was sold to Marvell who has been busy building
multi-core ARM-32 (and now ARM-64) parts which have 
taken over things like set-top boxes, small NAS boxes, etc.

SO now Intel is trying to play the TSMC game offering
merchant fab services! The only real takers seem to be
the FPGA guys who have no fear of spilling any beans
by having Intel fab their parts and they get a lot of
product lift because they can use an extra billion 
transistors. Intel has even done a couple of SoC parts
which put Xeon and a big&fast FPGA on the same
chip.

it's gonna be a long season for Intel as they try
to figure out life after the desktop PC.

   -mo