Good paper. But I must beg to differ on one claim in it (and will post
here so other's can counter-differ :-)
...The reason main memory is slow is economics, not because of some
fundamental limit -- DRAM is much cheaper than SRAM. Someday soon,
some company will figure out a cheap way to manufacture SRAM, and
the thousands of person-years spent finding clever ways to massage
the memory hierarchy will be suddenly irrelevant.
I believe the issue is fundamental physics and economics, and is never
going to go away. The issue has been around for at least 50 years now, and is
not DRAM vs. SRAM. The fundamental physics tells us that given a given
bits/volume density, the length of a path from a single CPU to N-bit
memory grows as the cube-root of N. The fundamental economics says that a
slow but much more dense memory technology will be attractive
to some users. For this reason, a memory hierarchy existed even for
very early machines. E.g. mercury delay lines versus vaccuum tubes.
In fact, even into the 60's (70's?) the largest (but slowest) memory
was the card decks in file cabinets.
Yes, someone will find a cheaper way to make SRAMs. But then someone
else will find an even cheaper way to make DRAMs (or some other
weird technology.)
This is not to contradict the main point of the paragraph:
Don't sweat too much about the memory hierarchy. The tuning issues
associated with memory hierarchies change yearly.
I agree with this, out of pessimism, not optimism about future hardware.
- Arch
--------------------- Object Oriented Numerics List --------------------------
* To subscribe/unsubscribe: use the handy web form at
http://oonumerics.org/oon/
* If this doesn't work, please send a note to owner-oon-list@oonumerics.org
This archive was generated by hypermail 2b29 : Wed Feb 20 2002 - 03:20:08 EST