Effect of model capacity on MLP versus gated-stride performance

Determine whether increasing the capacity of the multilayer perceptron changes its relative performance compared with the gated stride predictor on learnable memory-access traffic.

Background

The experiments use an MLP with only 257 parameters, and the paper notes that this limited capacity may contribute to its underperformance relative to the gated stride predictor on several learnable workloads.

Although the paper argues that the gate-closed safety property is independent of model capacity, it leaves unresolved whether a larger neural predictor would narrow or eliminate the observed performance gap against gated stride. This problem concerns relative predictor quality, not the structural safety theorem.

References

However, our finding---that the gate, not the predictor, produces the safety property---is capacity-independent; whether the relative performance of MLP vs. gated stride would change with a larger model remains open.

— Confidence-Gated Admission for Hardware Prefetching: When the Gate Matters More Than the Predictor  (2609.04040 - Majdane et al., 3 Sep 2026) in Section 9, Threats to Validity, Model capacity

The native ChampSim study validates the gate mechanism itself (gated vs. raw stride) but does not replicate the MLP-vs-stride matched control of \cref{tab:matched} externally; extending \cref{tab:matched}'s comparison to native SPEC CPU2017 traces is left to future work, and readers should not extrapolate the local matched-control finding to the native setting without that validation.

— Confidence-Gated Admission for Hardware Prefetching: When the Gate Matters More Than the Predictor  (2609.04040 - Majdane et al., 3 Sep 2026) in Section 9, Threats to Validity, External validity