2 ms·
This is not new, AMD has been using perceptron-based predictors probably even before Bulldozer. They are good, but harder to optimise for than standard rule bas
by AstralStorm 10y ago
This is not new, AMD has been using perceptron-based predictors probably even before Bulldozer. They are good, but harder to optimise for than standard rule based heuristic ones.
- sitkack 10y agoThey should expose the NN coefficients to the end user so that one can at least start from a known-better value. This problem would seem analogous to JIT warmup.