Physical substrates must earn their place, and this paper defines the entry fee

What the work claims

This is a position and analysis paper, not a primary experiment: no new devices, no new benchmarks, no measurements of its own. Its weight comes from synthesis and from a deliberately uncomfortable central proposition.¹ The claim is that photonic, in-memory and neuromorphic accelerators should be understood not as replacements for digital computing but as specialised engines hosted inside digital systems, and that a physical substrate earns an expanding role only where it delivers a measurable system-level advantage over a strong, optimized digital baseline for well-matched workloads, under digital orchestration that manages integration, uncertainty and fallback. To make that testable, the authors propose an eight-coordinate benchmark bundle that every serious physical-computing claim should report: task energy including interfaces, control and host involvement; end-to-end latency; task performance under declared conditions; volume and location of data movement; calibration overhead amortised over tasks; performance variation across temperature, supply, ageing and device variation; programmability and retargeting effort; and deployment cost including packaging, tooling, testing and certification. They also give the reason peak metrics fail: reported accelerator numbers are measured at incompatible boundaries. Their Table 1 makes the point with three examples: a 34-tile phase-change-memory chip in 14 nanometres reporting up to 12.4 tera-operations per second per watt chip-sustained but excluding the auxiliary digital compute and SRAM a product would need; a 64-core phase-change chip reporting 63.1 tera-operations per second and 9.76 TOPS/W at low precision but only 2.48 TOPS/W at high precision, chip-level including on-chip digital activation; and a photonic accelerator with over 16,000 components reporting up to gigahertz operation at potentially 3 nanoseconds per cycle, co-packaged with electronic logic, memory and control through 2.5D integration. Each number is real, the authors note, and none can be ranked against the others.

How it works

The paper's machinery is accounting, and it is built to be unforgiving. System energy is decomposed as sensing plus compute plus memory movement plus control plus radio plus actuation plus idle plus calibration. Two terms do most of the quiet damage to physical-computing optimism. Idle captures baseline and standby draw across the whole deployment interval, which is where always-on neuromorphic hardware actually wins and where exotic cores quietly bleed. Calibration captures programming, tuning, drift compensation and weight refresh, and the authors insist it be amortised over the number of tasks completed before recalibration is needed. A device that must be reprogrammed or re-tuned every few minutes carries that cost in every task it performs. On the suitability side they set four conditions for deployment: the workload must match the substrate, repeated operations must amortise programming and conversion costs, task accuracy must tolerate the substrate's effective precision or recover it cheaply, and the end-to-end advantage must survive after data movement, interfaces, host processing and communication are counted. They add a line that deserves quotation in every neuromorphic and biological-computing seminar: for event-driven systems, useful sparsity must exceed the baseline, state-maintenance and routing overhead, and, on neuron models, greater biological detail is not inherently preferable.

Where a skeptic should push

Because this is a position paper, the skepticism runs in both directions. Against the paper: it offers no new experimental evidence, its benchmark bundle is a proposal rather than a deployed standard, and an eight-coordinate non-scalar bundle, while honest, is also a recipe for incomparable reports unless the community agrees on workload, boundary and duty cycle alongside it. The authors themselves note the coordinates have different units and should not be collapsed into arbitrary weights, which is principled but leaves the door open to cherry-picked subsets. Against the field it criticizes: nothing. That is the point. The framework's most load-bearing assumption is that an optimized digital baseline is the correct reference, and on present evidence that assumption is sound, because digital systems keep absorbing the techniques physical computing champions, from sparsity exploitation to near-threshold operation, while retaining programmability, security and tooling. The demonstrated-versus-asserted split is clean: the incompatibility of reported measurement boundaries is demonstrated by Table 1; the claim that hybrid adoption will be incremental and workload-specific is a reasoned forecast, not a measurement.

The scorecard organoid computing must fill in

Living tissue is never mentioned in this paper, which is exactly why the paper functions as its most direct external audit. Run the organoid-computing program through the eight-coordinate bundle and the gaps are not subtle. Task energy with interfaces and host inside the boundary: essentially unreported anywhere; the incubator, perfusion, heated stage, gas control and amplifier chain are precisely the auxiliary digital-plus-analog overhead that Table 1 shows can erase a core's advantage. Calibration overhead amortised per task: for tissue this term is dominated by weeks of differentiation and maturation before the first query, plus per-batch characterization, and no published organoid-computing demonstration amortizes it honestly against task count. Field robustness across temperature, supply, ageing and variation: this is the most asymmetric coordinate, because biological batch-to-batch variability is large while the paper's own hybrid logic demands reproducibility or cheap digital correction. Deployment cost including tooling, testing and certification: uncharted for tissue, and the recent history of governance debate around brain organoids suggests certification will be a real cost, not a rounding error.

Now the opportunity, which is just as concrete. Two coordinates in the bundle are where a living substrate could genuinely beat every entry in Table 1. Calibration drift is the Achilles heel of analogue, photonic and memristive engines: conductances wander, phases drift, and the digital host must spend energy and time re-tuning them. Neural tissue is the one physical computing medium that calibrates itself, through homeostatic plasticity and activity-dependent adjustment, and that converts the calibration coordinate from a recurring tax into a substrate property. The idle coordinate is the other: a dormant organoid's metabolic draw is microwatts, and the paper's own logic says idle energy over the deployment interval is part of the bill. The catch, and it is the catch that decides the field's future, is that the tissue never deploys alone. Under this paper's accounting the incubator and interface are inside the boundary, so the self-calibrating advantage only survives if a future apparatus design drives the support overhead toward zero, for example with implantable or self-sustaining culture systems. The non-obvious implication is that the most valuable engineering contribution to organoid computing right now may not be neurobiology at all but apparatus minimalism, because every watt of support equipment is charged against every task forever.

The threat is a narrowing window. The paper's four deployment conditions read as a checklist for when organoids would be justified: sparse temporal processing, always-on sensing, continuous dynamics, and workloads that tolerate analog precision. Those are exactly the niches the field claims. Yet the same conditions are being met today by non-living neuromorphic and in-memory engines that already report wall-clock, wall-power numbers, however imperfectly bounded, and by hybrid systems in which digital orchestration supplies the reliability. And the paper's quietly devastating line, that greater biological detail is not inherently preferable, cuts at the methodological heart of organoid intelligence, which often treats biological realism as a virtue in itself. Under this framework realism is only worth what it measures. The dual-use and ethics notes are short but real: the paper frames digital orchestration as the layer that manages uncertainty, verification and fallback, which is also the layer where governance of a living compute substrate would have to live, and its insistence on reproducibility and certification raises the bar for any system whose substrate cannot be copied, reset or fully inspected.

The bottom line

Established by the paper's own assembly of cited results: accelerator efficiency claims are measured at incompatible boundaries and cannot be ranked, data movement and calibration are first-order energy terms, and a physical engine is justified only by end-to-end advantage over an optimized digital baseline. Asserted, reasonably but without new evidence: hybrid adoption will be incremental, workload-matched and digitally orchestrated. For organoid intelligence the paper is a free audit that arrived unsolicited: under its eight-coordinate bundle, today's demonstrations report perhaps two of eight coordinates, and the missing six are exactly where the living-substrate case is weakest. The calibrated verdict: the framework does not rule organoid computing out, it rules out the current standard of evidence. Confirming evidence would be an organoid system that reports the full bundle, including amortised maturation and support-apparatus energy, and still wins a coordinate against a digital baseline. Breaking evidence is accumulating by default, in the form of every non-living engine that reports more of the bundle each year.