Decompose the W4A8 accuracy gap caused by the full scaling contract

Determine the individual contributions of the full MAX7800x numerical contract’s scaling constraints to the W4A8 accuracy gap between ShiftQAT and less constrained quantization-aware training methods.

Background

The paper evaluates ShiftQAT under the MAX7800x accelerator’s full scaling contract, which requires uniform, symmetric, power-of-two, per-tensor scaling, signed 8-bit activations, static scale propagation, and hardware-compatible bias handling. At W4A8, ShiftQAT incurs larger accuracy losses than per-tensor QAT methods that relax some of these requirements, particularly asymmetric activations and learned non-power-of-two scales.

The paper identifies the remaining accuracy difference as the observed cost of enforcing the complete scaling contract at 4-bit weight precision, but does not isolate how much of the gap is attributable to each individual constraint or mechanism. Separating these effects would clarify which hardware restrictions most strongly limit low-bit accuracy and could guide improvements to contract-compliant training methods.

References

The remaining gap is therefore the observed price of the full scaling contract at 4 bits, and separating the contribution of each factor is left to future work (Section~\ref{sec:limits}).

— ShiftQAT: Quantization-Aware Training for the Numerical Contract of a Fixed-Function Edge CNN Accelerator  (2610.21695 - Buyuksolak et al., 15 Oct 2026) in Section 4.2, “Accuracy Under the Contract’s Scaling Constraints,” subsection “W4A8” (discussion following Table 2)