[Cryptography] Caches considered harmful

Kent Borg kentborg at borg.org
Thu Jan 11 10:11:22 EST 2018


One of the key properties of computers, one we take for granted and 
don't spend much time contemplating, is their predictability. We 
originally named them "computers" because their ability to compute--as 
in do arithmetic--was so impressive when compared to humans' ability to 
do the same.

With early computers there was no risk that they would replace humans in 
most endeavors because computers and humans are good at completely 
different things. But time passes.

We are now building and applying computers in mushy ways where they are 
no longer predictable, no longer deterministic. The two big examples of 
this, circa 2018, are (1) the magic tricks that are possible with neural 
networks and related self-tuning software, and (2) and all the 
self-tuning-usually-helps strategies that make a modern CPUs so much 
faster than a 68000.

Now, this isn't to say that the neural-inspired, speculative magic are 
all bad, just that we are trading away a key property of computers 
without making a conscious choice to do so: dogged predictability is a 
virtue. Seems we are trying to make them as fallible as we are, making 
them in our own image, without appreciating the implicit decision we are 
making.

-kb, the Kent who wants autopilots and MRI machines and cryptography 
programs and mortgage balances and door locks to be doggedly predictable.



More information about the cryptography mailing list