Championship Rules

Objective

This championship aims to foster innovation in branch prediction by encouraging the development of branch prediction algorithms that achieve high prediction accuracy while minimizing power consumption.

Eligibility

There are no eligibility requirements - as long as you have the motivation (whether you are passionate about branch prediction, or maybe just want a rematch) and are willing/able to make your submission following these rules and submission requirements, you are welcome to participate! You may submit as an individual or a team, affiliated with another organization or unaffiliated, etc.

Evaluation Criteria

Your submissions will primarily be awarded and selected for inclusion in the workshop based on how well your predictor maximizes the championship score (defined below) across the set of judging traces. The practicality of the design and quality of the paper and submitted code (including comments!) will also be considered. Novelty is encouraged, but will not be valued on its own merit (i.e. a novel design which performs worse than an old one is not better).

Scoring

The championship score will take into account prediction accuracy, energy-efficiency, and implementation complexity as part of a "Voltage-Frequency-Scaled Speedup" (a.k.a. VFS) score. Pierre has written a paper available on the championship's github repository explaining this scoring to help participants gain intuition for it.

Though we have attempted to ensure this scoring mechanism is aligned with real-world design constraints and you are heavily encouraged to use your creativity in maximizing it, submissions which optimize it in an unrealistic way or which undermine the intention of the scoring will not be considered. All submissions will be judged in the spirit of the championship: pushing the boundaries of energy-efficient and high-performance branch prediction.

Submission Requirements

1. All submissions will include (working) source code of the prediction algorithm, complete with comments which explain how it works and generally document the energy/area/latency costs.

2. You'll also submit a 4-page paper motivating your design, explaining how it works, assessing its power/area/timing costs, and analyzing simulation results. References and detailed addendum for cost analysis are not included in the 4-page limit. Additional information (including more detailed instructions for formatting and submitting your paper and code) is found on the submission information page.

3. All submitted algorithms must be implemented in the latest version of the championship model at the time of the submission deadline. You may not alter the model library code or the interface between the model and the prediction algorithm, but should implement your algorithm within it. The library does not allow the C++ code of prediction algorithms to read the values being operated upon at runtime (which is handled by the underlying hardware-modeling library) - your design must not subvert or otherwise work around this intentional constraint.

4. Your algorithm is not allowed to alter its behavior based on identification of individual traces. Modifying behavior in a generic way based on characteristics of the workload, however, is allowed and very much encouraged.

5. You agree that your submission (including but not limited to the paper, source code, and related information such as performance results and scoring) may be made public as part of this championship.