RE: software progress (RE: Hardware Progress: $319/GF)

From: Eugen Leitl (eugen@leitl.org)
Date: Sun Jun 02 2002 - 04:37:35 MDT


On Sat, 1 Jun 2002, Ben Goertzel wrote:

> Saying software tech hasn't progressed because it hasn't progressed in *your
> directions of choice* is like saying hardware tech hasn't progressed because
> it hasn't moved substantially toward massively parallel computing...

Actually, hardware is not doing well at all. Software and hardware are
locked in a tango, where no player can make significant advances on his
own. Progress, if/when it happens, is thus necessarily incremental.

All-purpose performance vs. transistor count has been lagging for a while.
We know that complexity is a dead end as high clock is concerned, yet the
CPU cores are getting more complex, not less. Asynchronous logic doesn't
really help here. Fraction of utilized CPU transistors for a task is
getting worse as Moore advances. There is no hardware message passing in
any mainstream CPU architecture. Because current software model is mired
in bloat, embedded DRAM has a hard time, since on-die grain size is
limited by yield reasons. Because of absence of embedded DRAM the nonburst
memory bandwidth increase is stagnating. Bad die yield goes up
exponentially with die size for a given process, but since our software
model doesn't do parallelism, there's pressure for as large dies as
economically possible. This makes wafer scale integration, which requires
semiquantitative yields (~90%) impossible. Reconfigurable logic is not
there yet. Runtime reconfigurable, adaptive logic has so far only been
prototyped for stochastic architectures. Cellular architectures (the only
way to go to high clocks and defect tolerance) are barely on the drawing
boards.

Given above toxic legacy, we're drawing orders of magnitude less
performance from a fixed amount of silicon real estate (say, a 300 mm
wafer in a modern process) than what would be possible in theory. Given
that premise, I'm distinctly underwhelmed by the claims, which just go GHz
MBytes magic mantra.



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT