• Metrics
• Charge

Charge to Metrics Group (DRAFT)

• Provide a recommendation for best practice on how we should quantitatively characterize the noisiness of a technology/environment and quantitatively judge the goodness/reliability of a component or system. Preferably, the system-level quantification is something that is useful in composition (knowing the metric on the subcomponents and some information about how subcomponents are composed, one can compute the metric on the larger ensemble).

Intended use

• specifying the challenge problems by the constituencies

• measure success of proposed solutions
• support CAD and automation
  • (if an optimization tool improves the goodness metric, we will agree it is doing the right thing)
• supporting division of labor and multi-layer coordination
  • (defining what needs to be communicated between device or circuit level and architecture/software level?)

Issues

• The broad range of noise challenges we address are more than just random bit-flips.
  • defects
  • variation
  • aging
  • hard failures
  • transients from ionizing particles, thermal noise, shot-noise...
    • Do we need to characterize region correlation (e.g. multi-bit impact of single strike)
  • This alone may demand the characterization of noisiness is a vector of numbers rather than a single scalar.
• Top-level result may need to be a vector as well
  • (SDC FIT rate, availability, ...)
• To support composition, perhaps component characterizations need additional information in the characterization vector?
• Are there open research issues in defining this that might need to be part of the recommended research program?

Strawman

• Characterize components by random upset rate (FIT rate)
• Compose using count of elements and some kind of structure/layer-vulnerability factor (derating, AVF)
• Measure system goodness with a metric like:
  • SDC FIT Rate
  • availability
  • survivability

Note

Here we're looking to understand the metric for reliability goodness. It should be understood that this could be used in making composite metrics that combine it with other physical goodness metrics like energy, area, and performance. e.g. maybe some people want to maximize reliability/energy or some product of reliability and performance.