The base of a split weight decides which noise you amplify

What the work claims

This is a numerical-analysis paper validated by simulation at three fidelities, not a silicon demonstration. The claim: when the dominant read noise of a crossbar tile is additive, meaning its magnitude does not scale with the stored code, then splitting each weight across two cells with a place-value base of b amplifies the high cell's noise by b squared in the summed output, and there is a closed-form base that minimizes the total error.¹ That base is b_opt = 1 + 1/(12 bound² eta²), where bound is half the cell's integer range (7 for a signed 4-bit cell) and eta is the per-tile output-noise standard deviation relative to the per-tile signal RMS, a single number measured once per device.¹ On a 512 by 512 device-physics simulation with programming noise, IR drop, and ADC quantization, the predicted interior optimum survives: mean per-layer MVM SNR peaks at b = 5, about 1.6 dB above equal-fill duplication and 3.3 to 3.4 dB above the common base-16 nibble split.¹ In end-application terms, at the same two-cell budget, the base split at b = 6 improves eight-task zero-shot transformer accuracy by 13.1 percent on Gemma-3 270M and 9.0 percent on Llama-3.2 1B relative to equal-fill, matching the strongest data-dependent baseline while using only uniform cell placement.¹

How it works

The encoding is a custom fixed-point format. Each weight is master-quantized onto an integer grid and carved into two integer cell codes: a high digit s_H and a low digit s_L, decoded as scale times (b s_H + s_L).¹ A base of 1 stores identical copies in both cells, pure noise averaging with no added resolution; a base of 2^q is non-overlapping bit-slicing, maximum resolution but the high slice's noise enters the sum weighted by a large place value; an intermediate base overlaps the two cells' ranges so the reconstruction grid stays gap-free while keeping the high cell's place value small.¹ The overlap constraint b at most bound + 1 keeps the low digit unsaturated; the naive base-16 split on 4-bit cells actually leaves holes in the representable grid, since values like 8 are unreachable when digits saturate at 7.¹

The error budget has two terms, and the paper's contribution is insisting on both. Read noise is additive per tile read, and because the periphery scales each tile's output by its place value before summing, the noise variance grows as (1 + b²): 37 for base 6 versus 257 for base 16, roughly a 7-fold difference.¹ Quantization is deterministic, a uniform grid of step set by the master scale, and its error falls as the grid gets finer with larger b.¹ Prior work modeled only the noise term, which is monotone in b, so its closed form lands on b = 1; adding the quantization floor bends the objective into an inverted U with an interior minimum.² The optimum is shallow: at eta = 0.02 the formula gives 5.25, both integer neighbors lie within 0.02 dB, and the whole band b = 3 to 8 is within 0.2 dB.¹ An optional fixed orthogonal rotation, applied to the weight matrix and undone on the input so the product is unchanged, Gaussianizes each tile's operands and is what makes one global base defensible: on Gemma 48 percent of tiles peak at b = 5 and 92 percent within {4,5,6}, with essentially no tile statistic predicting the per-tile best base.¹

Where a skeptic should push

The load-bearing assumption is additivity: per-tile output noise whose variance does not scale with the stored code. The authors are explicit that a conductance-proportional, multiplicative component pulls the optimum back toward b = 1, that real crossbars show state-dependent retention and programming error, and that the true device optimum lies somewhere between their additive interior base and the multiplicative duplication corner.¹ Which regime a given device sits in is an empirical question this paper answers only for its simulator, where the peak at b = 5 inverts through the closed form to eta near 0.02.¹ Note also what the device-physics figure shows: relative SNR gain over b = 1, with absolute levels omitted, and the end-application table is reported the same way, explicitly so that no absolute capability is disclosed.¹ That is a defensible trade-secret position for a memory vendor, but it means an outside reader cannot check how large the residual errors are in absolute terms, only that one encoding beats another on the same array. The evaluation is further bounded to a single matrix-vector product; error propagation through an iterative solver, where the per-step optimum need not be the fixed-point optimum, is left open.¹ And the attractive claim that one noise measurement replaces per-model search still requires that measurement to be made per device and workload, since eta aggregates programming noise, IR drop, ADC quantization, normalization, and operand distribution.¹

What this changes for MEA decode hardware

The reason this belongs on an MEA title is that the same crossbar fabric keeps appearing as the proposed back end for spike decoding: recorded neural data goes in as voltages, a trained decoder matrix sits in the cell conductances, and decoded output comes off the bit lines without the raw stream ever leaving the headstage. This paper says something non-obvious about that pipeline: the encoding of the decoder weights, not just the decoder accuracy in software, is set by the noise of the analog read chain, and that noise is the same quantity the acquisition front end spends its life fighting.

The mechanism is worth spelling out. Eta, the single dial in the closed form, is not a device-physics constant. The authors state it aggregates programming variation, IR drop, ADC quantization error, per-column normalization, and the operand distribution.¹ Every one of those terms is familiar to an MEA instrumentation engineer: IR drop is the wiring resistance between cells and sense amps, ADC quantization is the same converter resolution that sets the effective bit depth of the recording chain, and operand distribution is the statistics of the spike data being fed in. The practical consequence is that calibrating an analog decode fabric is not a separate exercise from qualifying the acquisition chain. A bench procedure that measures eta once per device, then evaluates the closed form instead of running a per-model numerical search, is cheaper than the evolutionary or empirical searches this result replaces,³ and it carries no programming or read cost over the default encoding.¹

The genuine opportunity is robustness. The optimum is a broad, flat basin: three to eight all within 0.2 dB for representative noise.¹ Analog arrays drift; a decoder encoding that stays near-optimal across miscalibration and device aging is worth more in a chronic implant or a long-running organoid rig than a peaky optimum that demands constant recalibration. The genuine threat is equally concrete. If a given fabric's residual noise turns out to be multiplicative, state-dependent, or dominated by retention drift, the promised 3.3 dB evaporates toward duplication, the same b = 1 optimum prior theory already prescribed.² The paper's own honesty about this is the caution: encoding recipes are regime-dependent, and a vendor datasheet that quotes cell precision without stating which noise regime the number was measured in is not telling you which split to use. For array architects the actionable habit is to ask for the measured noise-versus-code curve, not the nominal bit count, before believing any analog decode benchmark.¹

The bottom line

Established: under additive tile-output noise, which the paper's own device-physics simulation (including realistic non-additive IR drop and ADC effects) supports for its modeled array, the weight-splitting base has an interior optimum given in closed form by b_opt = 1 + 1/(12 bound² eta²), beating base-16 splitting by 3.3 to 3.4 dB in SNR and matching the best data-dependent encoding in transformer accuracy at uniform placement.¹ Asserted but unproven on silicon: that any specific commercial crossbar sits in the additive regime, and that absolute error levels after tuning are acceptable for neural decoding, since the paper deliberately withholds them. What would confirm the claim is the same measurement on a fabricated array: a noise-versus-code characterization showing a dominant additive component, plus a decaying multiplicative term that the interpolation extension can absorb.¹ What would break it is a device whose residual noise scales with stored conductance, which returns the optimum to duplication and this paper's headline gain to zero.