1. T-BAG: Bootstrap Aggregating the TAGE Predictor Ibrahim Burak Karsli, Resit Sendag University of Rhode Island

2. Bootstrap Aggregating Statistical method introduced by Breiman in 1996 Use ensemble of predictors –sub-predictors could be the same or different Train each slightly differently and independently Each predictor trained with resampled (with replacement) data set (bootstrapping) Aggregate their predictions The IDEA is: Many weak learners make strong learner Theoretically proven to perform better than single learner in an ensemble

3. Offline Bagging

4. Online Bagging

5. TAGE Predictor Winner of CBP3 State-of-art branch predictor Many parameters to allow variety

6. T-BAG: Prediction x32 TAGE PC aggregation prediction

7. Predictor Aggregation Bagging in nature uses 10s to 100s of predictors, so we target unlimited track Submitted predictor uses 32 TAGE predictors Keep track of successes of last 16 predictions with a sliding window for each predictor Aggregate the predictions using weighted sum

8. PC & resolveDir T-BAG: Update x32 TAGE Update Count

9. Random Update Each predictor is updated on each sample k times in a row where k is a random number generated by multinomial distribution Max k = 2 (because ctr width is 3bits) For submission, update on each sample 20%, 60%, 20% of the time, 0, 1, 2 times, respectively.

10. Sub Predictors 32 predictors Variability in min/max history lengths, number of tables, and use of PC in table indexing ctr 3-bits for all Each predictor’s size is about 15MB (submitted predictor 492MB) Min history varies between 3 and 13 Max history varies between 1,200 and 100,000 Number of tables varies between 20 and 38 16 predictors use PC, the other 16 do NOT! –Use of PC in indexing tables for TAGE-like predictor is not significantly better!

11. Results AllSame_RandUpd -> 1.952 misp/KI AllDifferent -> 1.932 misp/KI AllDisfferent_RandUpd -> 1.919 misp/KI

12. misp/KI ConfigurationBaseline TAGE32x-size TAGET-Bag32 AMEAN2.0072.0031.919

13. Conclusion and Future Work Simple idea Different types of predictors Implementation with storage budget

14. Q&A