Information-processing systems should be efficient, i.e., compute reliably while minimizing the consumption of resources. The efficiency of man-made information processors has been dominated by CMOS transistor scaling and Moore’s “law”. However, fundamental physical limitations are making it progressively more difficult for scaling to continue. As a result, innovative devices, circuits, and algorithms will be necessary for the computational revolution brought about by integrated circuits to continue. In this talk, I shall argue that millions of years of biological evolution provide us with an astonishing variety of potential solutions to this problem. Organisms face hard constraints of time, space, and energy, but must nevertheless function in a robust way within hostile environments in order to survive. As a result, they have evolved into extremely efficient information processors.
Biological inspiration can aid us at all levels of the information-processing hierarchy. I will discuss two fast-and-highly-parallel biologically-inspired mixed-signal chips as examples. The first example, an RF cochlea, maps the equations that describe fluid-membrane-hair-cell wave propagation in the biological cochlea into equivalent integrated circuits. I will show that the exponentially tapered transmission-line architecture of the mammalian cochlea performs constant-fractional-bandwidth spectrum analysis with O(N) expenditure of both analysis time and hardware, where N is the number of analyzed frequency bins. This is the best known performance of any spectrum-analysis architecture, including the constant-resolution Fast Fourier Transform (FFT), which scales as O(N log N). The RF cochlea uses this bio-inspired architecture to perform real-time, on-chip spectrum analysis at radio frequencies. I will demonstrate RF cochlea chips, implemented in standard 0.13μm CMOS technology, that decompose the RF spectrum from 600MHz to 8GHz into 50 log-spaced channels, consume <300mW of power, and possess 70dB of dynamic range. The real-time spectrum analysis capabilities of these chips make them uniquely suitable for ultra-broadband cognitive radio systems.
The second example exploits detailed similarities between the equations that describe chemical-reaction dynamics and those that describe subthreshold current flow in transistors to create fast-and-highly-parallel integrated-circuit analogs of protein and gene interaction networks inside a cell. Due to a natural mapping between the Poisson statistics of molecular flows in chemical reactions and electronic current flows in transistors, stochastic effects are automatically incorporated into my circuits, allowing computationally-intensive stochastic simulations of large-scale biochemical reaction networks to be performed rapidly. These chips carry out massively parallel computations, resulting in simulation times that are independent of model complexity. Currently, in spite of non-fundamental data-acquisition limitations, proof-of-concept designs simulate small-scale biochemical reaction networks at least 100 times faster than modern desktop computers. Much larger simulation-time speedups of genome and organ-scale intracellular and extracellular biochemical reaction networks are possible. By enabling improved analysis and design of such networks, these chips will be valuable tools for both systems and synthetic biology.