Automatic Tuning of Digital Circuits.

Abstract

Variation in transistors is increasing as process technology transistor dimensions shrink. Compounded with lowering supply voltage, this increased variation presents new challenges for the circuit designer. However, this variation also brings many new opportunities for the circuit designer to leverage as well.

We present a time-to-digital converter embedded inside a 64-bit processor core, for direct monitoring of on-chip critical paths. This path monitoring allows the processor to monitor process variation and run-time variations. By adjusting to both static and dynamic operating conditions the impact of variations can be reduced. The time-to-digital converter achieves high-resolution measurement in the picosecond range, due to self-calibration via a self-feedback mode. This system is implemented in 45nm silicon and measured silicon results are shown. We also examine techniques for enhanced variation-tolerance in subthreshold digital circuits, applying these to a high fan-in, self-timed transition detection circuit that, due to its self-timing, is able to fully compensate for the large variation in subthreshold.

In addition to mitigating variations we also leverage them for random number generation. We demonstrate that the randomness inherent in the oxide breakdown process can be extracted and applied for the specific applications of on-chip ID generation and on-chip true random number generation. By using dynamic automated self-calibrating algorithms that tune and control the on-chip circuitry, we are able to achieve extremely high-quality results. The two systems are implemented in 65 nm silicon. Measured results for the on-chip ID system, called OxID, show a high-degree of randomness and read-stability in the generated IDs, both primary prerequisites of a high-quality on-chip ID system. Measured results for the true random number generator, called OxiGen, show an exceptionally high degree of randomness, passing all fifteen NIST 800-22 tests for randomness with statistical significance and without the aid of a post-processor.