- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper co-designs a recurrent cell and an ultra-low-power analog circuit.
Critical strengths
- A genuine, verified algorithm–circuit correspondence. Every learned parameter has an explicit circuit counterpart: , , , and weights → width ratios. The paper validates this end-to-end at transistor level, not just block by block: per-stage overlays of candidates, states, skip outputs and logits (App. J), 49/50 HW–SW prediction agreement, Monte Carlo, and PVT corners. This depth of validation is uncommon in ML-venue hardware papers. It makes the "software model as a cheap hardware simulator" idea credible, at least at the validated scale, which is why I lean positive.
- A clean architectural insight. The BMRU state is binary with hysteresis and needs no settling dynamics, so it can be realized without capacitors. Its thresholding also confines analog errors to near-threshold events, instead of letting them integrate through time. The candidate→state error reduction in App. J (≈60 pA → ≈3 pA) illustrates the mechanism concretely, and the idea plausibly transfers to other analog inference platforms.
Critical weaknesses
- Hardware evidence covers a very small configuration, and the key scaling claims are extrapolated.
- Scale. All transistor-level evidence is at , two layers, binary KWS. Software accuracy there is 93.86%, below the 97–98% of the sub-µW digital 2-class KWS chips listed in Table 3.
- Extrapolation. Larger- power (Table 6), the noise study on sMNIST/pMNIST/dKWS (Fig. 3), and multi-class KWS all extrapolate from measurements. The quantities most likely to change with are exactly the ones being extrapolated. Current-summing nodes with fan-in accumulate mirror mismatch, so the "measured analog noise" from blocks need not hold at . Fig. 3 shows only a ≈2× margin before a sharp failure.
- Mismatch. At , the main-text example is misclassified in 11.5% of Monte Carlo instances. Monte Carlo covers only 11 clips, so the chip-to-chip accuracy distribution is unknown.
- Power budget. "Sub-µW up to " is not itemized. The extrapolated core is ≈640 nW at (Table 6). App. K says shift-register power (≈50 nW at ) scales linearly with the number of parameters, which grows roughly with . Several components are also excluded: input current generation for the MFCCs, majority-vote logic, and the front-end.
- Noise-suppression claim. "Suppress analog noise by at least 20-fold at each cell boundary" (abstract) generalizes from a mean absolute error on a single example. For binary states, the relevant metric is the HW–SW state-flip rate. The overlays show that flips and timing shifts do occur:
- Fig. 39, dim 4: a spurious HW pulse.
- Fig. 47, dim 3: different transition timing.
- Fig. 46, dim 3: the HW candidate is ≈0.7 where SW reaches ≈2.7. "Near-perfect agreement at every stage" is therefore overstated.
- Novelty and generality claims outrun what is shown.
- Scope of recurrence. The FQ reformulation is a modest change to an existing cell. Recurrence is per-unit (diagonal) with no recurrent weight matrix: the candidate depends only on the current input. Hence "linear marginal cost of recurrence" holds by construction. The noise-accumulation barrier is avoided by restricting recurrence to binary self-hysteresis, rather than overcome for general analog recurrent dynamics.
- "First fully analog RNN". This needs qualification. Analog and mixed-signal neuromorphic hardware with recurrent spiking connectivity, which also communicates through discrete events, has existed for years. The paper should state precisely which property is new.
- Expressivity argument. Prop. B.4 is an informal sketch ("except on a set of measure ε"), not a proof. The ListOps gap (37.5 vs 41.7) suggests that fixed thresholds do cost trainability.
- Table 1 backbone. Table 1 evaluates FQ BMRU inside a backbone with LayerNorm, GLU MLPs, sigmoid output gating and positional encoding, none of which the circuit can realize. It therefore does not characterize the deployable model.
Questions for authors
- Can you report the accuracy distribution across mismatch instances on the full test set? This could use Cadence, or the software proxy with the extracted mismatch statistics applied as static per-instance perturbations. Please do so at and . This directly affects how much the robustness claim can be trusted.
- How was the "measured analog noise" in Fig. 3 modelled: static per-network mismatch, per-timestep noise, or both? How does its magnitude scale with FC fan-in? Is the ≈2× margin expected to hold at ?
- Please give an itemized power budget at , covering core, mirror banks, shift registers, bias generation, input current generation and majority-vote logic. Is the total below 1 µW?
- Please report HW–SW binary state-flip rates per layer over the 50 clips, rather than only candidate/state MAE.
- Please state what is new relative to De Geeter et al. (2026) and Brandoit et al. (2026). In particular, did either already introduce unipolar outputs, fixed thresholds or a Schmitt-trigger realization? Please also report the hardware-backbone accuracies for the Table 1 tasks.
Update: 2 × 16 network validated at transistor level
Dear Reviewers,
We did not mention the 2 × 16 network below in our first responses because its simulations were still running. We chose to answer on time with the results we had, and to share these new ones as soon as they were ready.
All three reviews noted that the network validated at transistor level is small (2 × 4). We are happy to share that we have now simulated a 2 × 16 network in Cadence (5168 transistors). It uses the same topology as the 2 × 4 circuit, with the "yes" KWS model at 98.08 % in software (Table 4). To keep simulation time reasonable, we reused the 50 samples of Section 3.2, and the 11 Monte Carlo samples of Appendix H with 100 instances each.
In short: the 2 × 16 network behaves like the software model, it is more robust to mismatch than the 2 × 4 network, and its power matches our estimates. We attribute the better robustness to larger margins: the 2 × 16 network is more accurate, so its outputs sit further from the decision boundary. The software virtual chips show the same trend, with a smaller accuracy loss under mismatch at d = 16 than at d = 4.
Figures:
Proposed changes:
If these changes match your expectations, we will quickly upload a revised PDF. We plan to upload the revised PDF no later than Monday, November 16.
Best regards, The Authors