Refine FP8 Hybrid Predictor Features

Refine the Hamming-distance component of the Hybrid predictor for FP8 operands by incorporating weight-specific fields, such as mantissa-only toggles, to improve format-aware timing-criticality prediction.

Background

The Hybrid predictor combines activation switching information, represented by normalized Hamming distance, with an operand-magnitude feature. For FP8 operands, the magnitude feature is derived primarily from the exponent field, with optional leading mantissa bits. The current formulation computes Hamming distance over the full FP8 representation rather than isolating fields that may have different effects on multiplier delay.

The paper identifies field-specific Hamming-distance modeling—particularly tracking mantissa toggles—as an unresolved extension that could provide a more precise predictor for FP8 timing behavior.

References

For FP8 formats, HD is computed over the full 8-bit representation; refining HD to weight specific FP8 fields, such as mantissa toggles only, is left as future work.

— Dynamic Slack-Aware Clocking for Near-Threshold Tensor Processing Units (TPUs)  (2609.26644 - Nadeem et al., 22 Sep 2026) in Section 3.4(c), Hybrid Predictor