The traditional performance-cost benefits enjoyed for decades from
scaling of device area are challenged by two alarming trends: a
slowdown of both voltage scaling and frequency increase due to slower
scaling of leakage current as compared to area scaling, and a shift to
probabilistic design and less reliable silicon primitives due to
static and dynamic variations. These lead to a pessimistic projection that
it will be impossible to operate all on-chip resources, even at the
minimum voltage for safe operation, due to power constraints, and/or
the growing design and operational margins used to provide silicon
primitives with resiliency against variations will consume the scaling
benefits.
Our attempt, presented in this talk, towards reversing these negative
Trends is a first-order model that can determine analytically the
performance
degradation due to permanent faulty cells in prediction arrays. We
refer to this degradation as the performance vulnerability factor
(PVF). The model can predict, for a given program run, random
probability of permanent cell failure, and processor configuration,
the expected PVF as well as the PVF probability distribution for an
individual or combination of prediction arrays. This facilitates rapid
design space exploration of the performance implications of reliability
design decisions.