Predictive Design Space Exploration Using Genetically Programmed Response Surfaces Henry Cook Department of Electrical Engineering and Computer Science University of California, Berkeley [email protected]Kevin Skadron Department of Computer Science University of Virginia [email protected]
27
Embed
Predictive Design Space Exploration Using Genetically Programmed Response Surfaces Henry Cook Department of Electrical Engineering and Computer Science.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Predictive Design Space Exploration Using Genetically
Programmed Response Surfaces
Henry CookDepartment of Electrical Engineering and Computer Science
More transistors = more complex designs Superscalar, out-of-order execution Chip multiprocessors, systems-on-a-chip Heterogeneity, shared memory, interconnects…
Many interrelated design variables with emergent interactions Unclear whether analytic models can be
constructed based on a priori understanding
How can we find the best design?
Detailed, cycle-accurate simulations Testing every option is time consuming
1 min. x 1 billion simulations = too long! Even search requires too many
simulations Solution: Predict which designs are best
based on a small data sample
Why not heuristic search?
Every step of search requires at least one full simulationHow many steps?
Predicting based on a predetermined sample becomes increasingly efficient as space growsAssuming predictions are accurateAssuming sample is small enough
How can we address complexity?
We need a tool to predict global performance based on a data sample
Many techniques might do this, but we want to use one that… Is more accurate when using small samples Finds the true optimal solutions Gives the architect more insight
Not just ‘what’ the best answer is, but also ‘why’ it is the best
How can we make predictions?
Build a ‘response surface’ A function that relates
design choices to a performance measure
Predicts performance for designs we did not ever simulate