Principles of Top-Down Mixed-Signal Design
Consider a digital design. In most cases digital systems are implemented as finite-state machines (FSM) and constructed from standard cell libraries. Using a FSM formulation acts to unify and homogenize digital design and gives it a well-understood mathematical foundation. This foundation was thoroughly explored in the late ’80s resulting in the commercial logic synthesis tools of the early ’90s. These tools take RTL, a relatively high-level description of a digital system that is created by designers and can be verified with the help of logic simulators, to produce an optimized gate-level description of the system. This transformation is possible because digital systems are constructed from a limited set of relatively simple and well-behaved building blocks. The building blocks of digital systems are gates and registers. The blocks, generally referred to as cells, all share a common I/O model and so are easily interconnected, are derived from a rela- tively small number of cell types that have very simple and easily described behavior, are easily parameterized in terms of the number of inputs and outputs, and have a simple and easily adjusted performance trade-off that involves only speed and power. Logic synthesizers operate by creating a complete mathematical description upon which it per- forms transformations in order to create an optimal design in terms of speed, power, and area. This is a two step process. First, equivalence transformations are applied to the mathematical descriptions in order to reduce the total number of gates, which minimizes the area, and the depth of the logic, which roughly maximizes the speed. This is possible because each block has a simple logical description and a common interface model. Then, the drive ability of each gate is adjusted to provide the lowest power while still meeting speed requirements. This is possible because this speed-power trade-off is eas- ily made in each gate.
Now consider analog design. Analog design has no equivalent to finite-state machines, and so has no unified formulation and no common mathematical foundation. It also has no universal equivalence transformations that allow the topology of a circuit to be easily modified without risk of breaking the circuit. These problems prevent a topological mapping from a behavioral description to hardware. Even if one were mathematically possible, the lack of a common I/O model for analog blocks would prevent the topolog- ical modifications that are needed for either mapping or topology optimization.
It might be possible to try to enforce a common I/O model for analog circuits, but doing so would be very expensive. For example, one might simply specify that the signals at the inputs and outputs of analog blocks center around a particular value, have the same maximum swing, and that outputs have zero output impedance and inputs have zero input admittance. The problem is that doing so would necessitate extra circuitry in each analog block that is there simply to assure compliance to the I/O model. That circuitry reduces the overall performance of the circuit by increasing power dissipation, increas- ing noise, decreasing bandwidth, etc. This differs from the digital world where the com- mon I/O model was achieved naturally and without significant cost. In addition, it is not possible to achieve these ideals at high frequencies. Instead, some common reference impedance would have to be specified, such the 50Ω used at the system level, but driv- ing such loads greatly increases power dissipation.
Finally, there is no simple way to trade-off the various performance metrics that are important with analog blocks, which makes it very difficult to perform a parametric optimization. Sensitivity-based local optimizations can be used, but the improvement provided by these approaches is usually small. Monte Carlo-based global optimizers offer better improvements, but require substantially more in the way of computer resources.
The Designer’s Guide Community
3 of 31