Information Technology Reference
In-Depth Information
where w is the multiplicative factor, and M is the number of months between initial
and final estimates.
Y ¼ M
=
R
ð 17
:
2 Þ
so
w M
¼ 2 ð M = R Þ
ð 17
:
3 Þ
where Y is the number of times the transistor count doubles, and R is the number
of months elapsed for the transistor count to double. Assuming doubling in
18 months, and assuming we look 180 months (15 years) into the future from
2006, to the year 2021, the transistor counts should have doubled 10 times.
w 180
¼ 2 ð 180 = 18 Þ ¼ 2 10
¼ 1024
ð 17
:
4 Þ
c fin ¼ c init w 180
¼ 10 12 transistors
=
chip
ð 17
:
5 Þ
Therefore, we predict that in 2021 there would be 10 12 transistors/chip.
We assume that on the average, each cortical neuron has 10,000 synapses [9].
If each synapse contains 10 transistors, plus the few transistors representing the
axon, the total number of transistors in a neuron is 10 5 . Therefore, the number of
neurons per chip can be estimated to be 10 7 .
Thus, in 15 years, we will require 10 4 chips (integrated circuits) to construct an
artificial brain with 100 billion neurons. Boahen predicts biomimetic chip densities
will be within a factor of 10 of biological neuron density within the decade [1];
however, his estimates regarding the number of transistors per synapse and
number of distinct synapses differ from ours.
There is more uncertainty when estimating system size because the future of
multi-chip modules (MCMs) is less predictable. However, we will assume MCMs
occupy 30% of board space. On each MCM, we assume dies occupy 70% of the
space. We assume that six 12 00 12 00 boards fit vertically in a cubic foot of space.
Then the system could hold 180 chips/ft 3 . Based on these assumptions, in the year
2021, we require a space of 55.5 cubic feet to house our neural circuits, absent any
room required for interconnections between neurons on chip, between boards and
between racks. If the racks are 8 feet tall, then we require 6.9 square feet of
floor space for the equipment. Allowing for air space around the equipment,
we estimate the neurons in the artificial brain to occupy 14 sq. ft. of floor space.
This is approximately the total free space available in the first author's university
office.
This can be contrasted with the IBM/Swiss project that allocates only two
neurons per processor, highlighting the economies of a hardware solution.
Neurons with learning capabilities are significantly larger than the ones on which
 
Search WWH ::




Custom Search