From: Eugen Leitl (eugen@leitl.org)
Date: Sat Jun 01 2002 - 05:17:59 MDT
On Fri, 31 May 2002, Dani Eder wrote:
> The cost of processing power continues to drop like
> a rock with rocket assist. Cost has dropped 20%
> in the past two months and approximately in half
> in the past 6 months.
How does memory bandwidth fare, in comparison? You should plot that, and
not just stream data.
Integration density is real. However, CPU architecture moves towards more,
nor less complexity. This implies higher setup latency and less
general-purpose suitability, plus sets hard limits on clock scalability in
a given process. Software progress is stagnating, and the current software
model is locking in current hardware model. Single-point breakthrough
won't translate into traction immediately.
> If optimistic estimates of the required computer
> power for human-level AI are correct at 100 TFlop/s,
> it presently costs $31.9M to buy a human's worth
> of computers.
COTS system installations of 31.9 M$ are cost dominated by air
conditioning, power, and human administrators. Hardware failure rate will
be unacceptably high for brittle systems. Network interconnect will be
your bottleneck -- assuming your code parallelizes nicely. Embarrassingly
parallel codes are exempt from this, obviously.
> Economic crossover is not implying we would have
> human-level AI at that point, but rather that
> skilled human tasks which require lots of processing
> power (i.e. driving a vehicle on a highway) would
Would have to be a small box, and low-power at that. Would profit from
novel sensorics (ToF radar) and dedicated hardware (silicon retina).
High threshold: liability issue.
> be cost-feasible. This has implications in the
> social and economic realms.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT