LAUSR

Intrinsic resiliency of many today's applications opens new design opportunities. Some computation accuracy loss within the so-called resilient kernels does not affect the global quality of results. This has led the scientific community to introduce the approximate computing paradigm that exploits such a concept to boost computing system performances. By applying approximation to different layers, it is possible to design more efficient systems—in terms of energy, area, and performance—at the cost of a slight accuracy loss. In particular, at hardware level, this led to approximate integrated circuits. From the test perspective, this particular class of integrated circuits leads to new challenges. On the other hand, it also offers the opportunity of relaxing test constraints at the cost of a careful selection of so-called approximation-redundant faults. Such faults are classified as tolerable because of the slight introduced error. It follows that improvements in yield and test-cost reduction can be achieved. Nevertheless, conventional automatic test pattern generation (ATPG) algorithms, when not aware of the introduced approximation, generate test vectors covering approximation-redundant faults, thus reducing the yield gain. In this work, we show experimental evidence of such problem and present a novel ATPG technique to deal with it. Then, we extensively evaluate the proposed technique, and show that we are able to achieve an average yield improvement ranging from 19% up to 36% — compared to conventional ATPG—in terms of approximation-redundant fault coverage reduction. In some cases, the improvement can reach up to 100%.