A High-Precision Current-Mode WTA-MAX Circuit with Multi-Chip Capability

Abstract

Abstract-This paper presents a circuit design technique suitable for the realization of winner-take-all (WTA), maximum (MAX), looser-take-all (LTA), and minimum (MIN) circuits. The technique presented is based on current replication and comparison. Traditional techniques rely on the matching of an N transistors array, where N is the number of system inputs. This implies that when N increases, as the size of the circuit and the distance between transistors will also increase, transistor matching degradation and loss of precision in the overall system performance will result. Furthermore, when multichip systems are required, the transistor matching is even worse and performance is drastically degraded. The technique presented in this paper does not rely on the proper matching of N transistors, but on the precise replication and comparison of currents. This can be performed by current mirrors with a limited number of outputs.Thus, N can increase without degrading the precision, even if the system is distributed among several chips. Also, the different chips constituting the system can be of different foundries without degrading the overall system precision. Experimental results that attest these facts are presented.

An analog scheme for fixed-point computation-Part II: Applications

“…This minimum circuit uses much less die area and is typically faster and more accurate than a minimum circuit comprised of a tree of voltage comparators [23]. Recent developments indicate that, using alternative topologies, it is possible to build well-matched maximum circuits that span several die, thus increasing the number of nodes that can be processed reliably [31]. The circuit settles to the maximum of the input currents provided the dc loop gain around each pair of current mirrors is 1.…”

Adaptive Resonance Theory Microchips

“…3.3, all M A transistors must match well, which is very unlikely if they are in different chips. A solution for this problem is to use a WTA circuit technique based on current processing and replication, insensitive to inter-chip transistor mismatches [Serrano-Gotarredona, 1995], [Serrano-Gotarredona, 1998a], which is described in Chapter 4.…”