{"id":"1638d1bd-743f-42e8-a489-0d5969c78014","arxiv_id":"2504.17752","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A passive RF mixer and OFDM-style frequency encoding compute fully connected layers over the air, reporting 95.7% MNIST accuracy and a modeled 6.0 fJ/MAC client energy.","lead":"WISE offloads deep learning matrix math onto radio frequency signals: a central radio broadcasts model weights and a passive mixer on each client multiplies them with local data in the analog domain. The paper reports 95.7% MNIST accuracy and a computed 6.0 fJ/MAC client-side energy, but the energy number comes from an analytical model, not a measured power draw.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy-efficiency claim rests on a client-only component model that excludes the central-radio broadcast power driving the passive mixer; at the demonstrated scale it is not a measured operation power.","rationale":"The paper's internal convolution/MVM math is consistent, and the 95.7% accuracy at 25 dB SNR is a real experimental observation. The load-bearing weakness is the energy metric attached to that accuracy. The abstract and Results explicitly use 6.0 fJ/MAC and 165.8 TOPS/W as the demonstration of 'ultra-low operation power' and as the basis for the two-to-three-orders-of-magnitude comparison with digital ASICs. But Eq. (2)/S105 is an analytical model of selected client components, not an energy measurement, and it omits the physical source of the mixer's LO drive. Because the LO drive is not a negligible overhead—it is the central broadcast itself—the metric as stated cannot be verified from the data provided. The reader's conditional verdict is appropriate: the analog computing concept and accuracy results are credible, but the headline energy-efficiency claim needs either a system-level measurement including broadcast and SDR power or an explicit, clearly labeled redefinition as a client-ASIC lower bound with externally supplied LO. I agree with the reader's identification of the energy accounting as the weakest assumption.","tokens_in":54405,"tokens_out":11079,"duration_ms":116681,"concrete_test":"On the existing SDR testbed, measure the DC wall-plug energy of the central radio and of one WISE-R over a full MNIST inference at 25 dB SNR with the broadcast active. Compute system energy per client-MAC as (E_central/3 + E_client)/1.06e6. Also repeat with the WISE-R LO port driven by a local signal generator instead of the broadcast W to isolate the LO/broadcast contribution. If the measured per-client energy including central broadcast exceeds the modeled 6 fJ/MAC by more than 10x, the headline energy claim must be restated as a component-level lower bound and the 165.8 TOPS/W cannot be claimed as demonstrated operation power.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (2) / Supp. Eq. (S105) defines the headline 6.0 fJ/MAC as E1+E2+E3, where E1 is client TX energy for x(t), E2 is ADC sample energy, and E3 is digital FFT energy. This model assumes the mixer's LO drive is free. In the experiments, the LO port of the ZEM-4300+ is driven by the wirelessly received weight waveform at -3 dBm, supplied by the central radio's broadcast; Supp. §12C reports a central average TX power of about 9 dBm. For the MNIST demo, the broadcast lasts about 17.7 ms per inference, so the central radio consumes about 140 uJ per inference, or about 47 uJ per client across the three clients. Per real-valued MAC (1.06M per inference), that is about 44 pJ/MAC, roughly four orders of magnitude above the reported 6 fJ/MAC. Adding the client-side model gives about 50 pJ/MAC end-to-end, or about 0.02 TOPS/W, not 165.8 TOPS/W. The paper also uses literature values eadc = 1 pJ/sample and edig = 1 pJ/MAC rather than a wall-plug measurement of the SDR. The reported numbers are therefore a lower-bound design estimate for an idealized WISE-R ASIC powered by an external LO/broadcast, not an operation power demonstrated on the SDR testbed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WISE, an edge-computing architecture in which a central radio broadcasts frequency-encoded neural-network weights over the air and each client uses a passive RF mixer to compute the matrix-vector product y = Wx in the analog domain. The authors derive an OFDM-based frequency mapping that realizes the MVM as a convolution during frequency mixing, implement a three-client USRP X310 plus Mini-Circuits ZEM-4300+ testbed, and report inner-product RMSE and MNIST/AudioMNIST classification accuracy as functions of SNR. They report 95.7% MNIST accuracy at 25 dB SNR and claim an energy efficiency of 6.0 fJ/MAC, corresponding to 165.8 TOPS/W, computed from an energy model that combines thermal-noise-limited client waveform generation, ADC sampling, and digital FFT decoding.","tokens_in":54695,"tokens_out":11936,"duration_ms":119334,"significance":"The experimental MVM and classification results are externally benchmarked and are not fitted to the energy model; the accuracy degrades with SNR as expected, which lends credibility to the in-physics MVM mechanism. The disaggregated broadcast idea is a worthwhile architectural contribution, and the supplementary material is unusually detailed about the experimental setup. However, the headline energy-efficiency claim is not a measured operation power: it is assembled from literature ADC/ASIC energies and an explicitly client-only model that excludes the central broadcast energy. For the demonstrated three-client setup, the broadcast energy dominates by roughly four orders of magnitude, so the system-level efficiency gain over digital ASICs is not established and the claimed two-to-three order improvement should be substantially qualified.","major_comments":[{"comment":"The headline claim of 6.0 fJ/MAC and 165.8 TOPS/W is not a measured operation power. Eq. (2) (and Supp. Eq. S105) computes e_mvm from a component model that assumes e_adc = 1 pJ/sample and e_dig = 1 pJ/MAC, and Supplementary Section 7B explicitly excludes the central radio's broadcast energy. In the demonstrated MNIST setup the central radio transmits at about 9 dBm average power for roughly 17.7 ms per inference (Supplementary Sections 12C and 15), consuming about 140 uJ; with three clients this is about 44 pJ per client per real-valued MAC, four orders of magnitude larger than the reported 6 fJ/MAC. The reported numbers are therefore a lower-bound design estimate for an idealized WISE-R ASIC, not the operation power of the demonstrated SDR system. The abstract and Results should be revised to state this, and a system-level energy accounting should be provided as a function of the number of clients, together with the number of clients at which the per-client broadcast cost becomes negligible.","section":"Methods Eq. (2); Supplementary Sections 7B, 12C, 15"},{"comment":"The 'thermodynamic limit' e_tdl is defined as the N-to-infinity limit of the same energy model in Eq. (2), not as an independently derived physical bound. Claims that WISE 'approaches the thermodynamic limit' and 'surpasses the Landauer bound' are therefore statements about the model's own asymptotic, and they inherit all omissions of the model, including the broadcast energy and the assumed ADC/ASIC energies. This should be relabeled as a model-internal asymptotic limit, and the comparison with Landauer should either be derived from a first-principles accounting of the actual physical resources or removed.","section":"Methods Eq. (3); Supplementary Section 8F"},{"comment":"The experimentally reported accuracy values are accompanied by energy-efficiency numbers evaluated from Eq. (2), not from any power measurement of the SDR testbed. Because e3, the digital FFT decoding term at an assumed 1 pJ/MAC, is a substantial fraction of the reported 6.0 fJ/MAC for MNIST, the phrase 'experimental energy efficiency' in the Results section is misleading. The actual USRP X310 power draw is not reported, and the SDR itself does not operate at fJ/MAC scale; the values should be called projected efficiencies from the component model, with all assumed parameters stated in the main text.","section":"Results, Fig. 4b; Methods, Eq. (2)"}],"minor_comments":[{"comment":"The abstract says 'ultra-low operation power of 6.0 fJ/MAC'; fJ/MAC is energy per operation, not power. Please use 'operation energy' or 'energy per MAC' consistently.","section":"Abstract"},{"comment":"The Fig. 2e caption reports MNIST accuracies of 97.1% to 97.4% across the three clients, while the main text reports 95.7% at 25 dB SNR as the headline accuracy. The caption appears to correspond to the x-precoding scheme in Supplementary Section 13 rather than the W-precoding configuration used for the headline number; please reconcile this inconsistency.","section":"Fig. 2e caption"},{"comment":"The Landauer-limit expression appears garbled: 'eLandauer = b2· ln 2·kT0' should be a clearly defined formula in terms of bit precision b. Please correct the typography and define the comparison precisely.","section":"Supplementary, Eq. (S87)"},{"comment":"The caption contains a typo: 'Mino-Circuits ZEM-4300+' should be 'Mini-Circuits ZEM-4300+'.","section":"Fig. S6 caption"},{"comment":"The computation throughput is reported in OPS, but the quantity is real-valued MACs per second; please define OPS explicitly in the main text to avoid confusion with floating-point operations.","section":"Methods, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read this one. First, the experimental core is real: a central radio broadcasts OFDM-encoded model weights over a 25 MHz wireless link, three USRP clients mix those weights with locally generated input waveforms using a passive diode mixer, and the resulting frequency-domain convolution gives believable MVM results. MNIST accuracy around 95.7% at 25 dB SNR, with the expected accuracy-versus-SNR behavior, is a legitimate small-scale proof of concept. Second, the headline 6.0 fJ/MAC and 165.8 TOPS/W are not measured operating numbers. They come from Eq. (2), which sums client TX energy, ADC sampling energy at an assumed 1 pJ/sample, and digital FFT energy at an assumed 1 pJ/MAC. That model excludes the central radio's broadcast energy and treats the mixer's LO drive as free.\n\nWhat is actually new: the specific combination of W-precoding for channel compensation, frequency-encoded broadcast weights, time-domain encoding of x on the client, and MVM decomposition into small FFTs. The convolution theorem trick is known, and AirNN/AirFC and photonic frequency-domain MVM cover pieces of it, but I do not see an earlier implementation of this exact passive-mixer client. The supplementary is thorough and the math is internally consistent. The paper also deserves credit for explicitly stating in Supp. 7B that client energy excludes the central radio broadcast. That is an honest limitation, but it is fatal to the abstract's phrasing.\n\nThe load-bearing soft spot is the energy accounting. In the experiments, the LO port of the ZEM-4300+ is driven by the wirelessly received weight waveform at about -3 dBm, and the central radio broadcasts at roughly 9 dBm average power. For an MNIST inference the broadcast lasts on the order of 17.7 ms, which is about 140 uJ of central transmit energy, or about 47 uJ per client across three clients. Spread over 1.06M real MACs, that is about 44 pJ/MAC before adding the client-side model, roughly four orders of magnitude above 6 fJ/MAC. Even with the client energy terms included, end-to-end efficiency is on the order of 0.02 TOPS/W, not 165.8. If the claim is reframed as 'client-side MVM energy under an idealized ASIC WISE-R powered by an external LO/broadcast,' the numbers are a defensible design estimate. As stated, they overstate what was demonstrated by a wide margin.\n\nOther issues are minor by comparison. No code or data is released, no error bars are given, and the model assumptions e_adc and e_dig are literature values rather than wall-plug measurements. The novelty is moderate but the paper is honest about prior over-the-air computing work.\n\nWho is this for? People working on analog or wireless computing, and reviewers who care about what 'operation power' means in an accelerator paper. It deserves a serious referee, but the referee should ask for either a system-level energy measurement or a clearly labeled lower-bound estimate with the broadcast power included as a separate curve. My recommendation: send it to peer review, but treat the energy headline as unproven until the accounting is fixed.","headline":"The SDR demonstration is real and the math holds up, but the 6.0 fJ/MAC headline is a client-only component estimate that excludes the central broadcast power actually driving the mixer, so the energy claim is not demonstrated as stated.","tokens_in":55300,"tokens_out":2188,"would_cite":true,"duration_ms":24904,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Neural-network math falls to 6 femtojoules per MAC via a radio mixer.","keywords":["in-physics computing","radio-frequency computing","over-the-air model broadcast","passive frequency mixer","energy-efficient deep learning inference","disaggregated edge computing","matrix-vector multiplication"],"falsifier":"Measure the end-to-end energy of one WISE-R inference including the central radio's transmit power and the client's local-oscillator source, then divide by the number of real-valued MACs; if the total exceeds the quoted 6.0 fJ/MAC by orders of magnitude, the operational-power claim fails. A second check: replace the diode mixer with an ideal analog multiplier and repeat the N=4,096 inner-product test; 5-bit accuracy should improve in the high-SNR regime if thermal noise, not mixer switching, is the floor.","tokens_in":54161,"feed_emoji":"📡","tokens_out":11883,"duration_ms":97214,"temperature":0.7,"pith_summary":"WISE proposes that deep-learning inference on edge devices does not need the model stored locally or multiplied digitally: a central radio broadcasts the weights as a radio waveform, and each client multiplies that waveform with its own data in a passive frequency mixer, so the matrix-vector product $y = Wx$ emerges from the physics of frequency mixing. The paper argues that this disaggregated, in-physics design removes the dominant memory-access and multiply-accumulate costs from the client, leaving energy dominated by waveform generation, sampling, and decoding. On a software-defined radio testbed with over-the-air broadcasts, WISE reports 95.7% MNIST classification accuracy at 6.0 fJ per real-valued MAC (165.8 TOPS/W), and argues that per-MAC energy improves as the problem size grows, approaching a thermodynamic limit set by thermal noise. Battery-powered cameras, drones, and IoT nodes could then run useful models without storing them, at energy costs two to three orders of magnitude below digital ASICs.","feed_headline":"6 femtojoules per MAC: a passive radio mixer computes neural networks","feed_subtitle":"Broadcast weights meet local data in a passive mixer, cutting edge-device inference energy by orders of magnitude.","key_machinery":"The load-bearing object is the hybrid convolution theorem for OFDM waveforms: when two time-domain OFDM waveforms are multiplied, their frequency-domain symbol spectra linearly convolve; with W's entries placed on subcarriers and x's entries on every M-th subcarrier, the middle M subcarriers of the convolution equal $y = Wx$. A passive double-balanced diode mixer performs the time-domain multiplication, and channel state information is folded into a precoder $V = W/H$ at the central radio so the client receives undistorted weights. A Zadoff-Chu phase sequence in the activation function spreads the power of the next layer's waveform evenly across the spectrum.","core_discovery":"The central claim is that a general complex-valued matrix-vector multiplication $y = Wx$ can be carried out in the analog radio-frequency domain. W and x are mapped onto orthogonal subcarriers of two OFDM waveforms; a passive mixer multiplies the time-domain waveforms, and the hybrid convolution theorem makes the spectrum of the product contain the entries of $Wx$ in a narrow band, which the client samples and decodes. The paper reports that this in-physics MVM achieves over 5-bit computing accuracy for inner products up to $N=32{,}768$, and that on MNIST and AudioMNIST the three-client WISE-R implementation reaches 95.7% and 97.2% classification accuracy at 6.0 fJ/MAC and 2.8 fJ/MAC, corresponding to 165.8 and 359.7 TOPS/W. The paper further claims that as $N$ grows, per-MAC energy approaches $e_{\\mathrm{tdl}} = \\mathrm{SNR}\\, k_B T_0 / 4$, the thermal-noise limit of analog hardware, which can sit below the Landauer bound for the same bit accuracy.","pith_inferences":["The paper's headline energy accounting counts only the client side; a system-level fJ/MAC that also charges the central radio's broadcast power, the mixer's local-oscillator drive, and the activation functions will be higher unless those costs are amortized over many clients and inferences.","The same frequency-mixer mechanism could be applied to other analog inner-product workloads, such as fixed-code correlation or filter-bank processing, where the broadcast waveform is a known kernel rather than a trained weight matrix.","A clean hardware check of the accuracy model is to replace the diode mixer with an ideal analog multiplier: if 5-bit accuracy does not improve in the high-SNR regime, then mixer switching, not thermal noise, is the actual error floor."],"forward_implications":["A client can run a three-layer fully connected model for image or audio classification without storing the model, receiving freshly broadcast weights on demand.","Per-client energy per MAC drops as the layer's input dimension N grows, because ADC sampling and FFT decoding costs amortize; the paper measures 2.4 fJ/MAC at N=4,096 and 1.4 fJ/MAC at N=32,768 for inner products.","Computation throughput across U clients grows as $4UB/[(1+\\alpha)(1+\\beta)]$, so more bandwidth or more clients directly buys more MACs per second.","With ideal hardware the per-MAC energy approaches $\\mathrm{SNR}\\, k_B T_0/4$, a thermal-noise floor that can fall below the Landauer limit for 4-bit and 5-bit accuracy.","The same broadcast-and-precode structure extends in the paper's analysis to other MVM-based models, including convolutional neural networks and transformers."],"supporting_citations":[{"why":"Supplies the 1 pJ/MAC digital-ASIC baseline against which WISE's fJ/MAC and TOPS/W claims are measured.","marker":"[7]"},{"why":"Provides the LeNet-300-100 architecture and the MNIST dataset used as the main classification benchmark.","marker":"[33]"},{"why":"Defines the Zadoff-Chu sequence used in the activation function to spread waveform power across subcarriers.","marker":"[34]"},{"why":"Is the passive double-balanced mixer hardware that performs the in-physics multiplication in the experiments.","marker":"[49]"},{"why":"Supplies the 1 pJ/sample ADC energy assumed in the energy-efficiency model.","marker":"[52]"},{"why":"Defines the Landauer limit that WISE's thermodynamic limit is compared against.","marker":"[9]"},{"why":"Provides the large-model dimension (N=11,008 for Llama-2-7b) used to argue that per-MAC energy improves at scale.","marker":"[6]"},{"why":"Supplies the AudioMNIST spoken-digit dataset used for the second classification task.","marker":"[41]"}],"fun_headline_variants":["6 fJ/MAC neural nets via passive RF mixing","In-physics RF computing hits 165.8 TOPS/W on edge","Broadcast weights, mix passively: deep learning at 6 fJ/MAC","Radio waves compute matrix math: 100x less energy for AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline energy numbers stand or fall on whether a client's energy really is just the reported waveform-generation, sampling, and decoding costs, with the broadcast, mixer drive, and activation costs negligible once amortized.","fun_headline_variants_meta":{"raw":{"variants":["6 fJ/MAC neural nets via passive RF mixing","In-physics RF computing hits 165.8 TOPS/W on edge","Broadcast weights, mix passively: deep learning at 6 fJ/MAC","Radio waves compute matrix math: 100x less energy for AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000367,"raw_usage":{"total_tokens":1992,"prompt_tokens":987,"completion_tokens":1005,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":927}},"tokens_in":603,"tokens_out":1005,"duration_ms":9722,"temperature":1.0,"reasoning_tokens":927,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:33:24.998699+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the end-to-end energy of one WISE-R inference including the central radio's transmit power and the client's local-oscillator source, then divide by the number of real-valued MACs; if the total exceeds the quoted 6.0 fJ/MAC by orders of magnitude, the operational-power claim fails. A second check: replace the diode mixer with an ideal analog multiplier and repeat the N=4,096 inner-product test; 5-bit accuracy should improve in the high-SNR regime if thermal noise, not mixer switching, is the floor.","supporting_citations":[{"cited_title":"Horowitz, Computing’s energy problem (and what we can do about it) , in Proc","cited_arxiv_id":null,"evidence_quote":"Supplies the 1 pJ/MAC digital-ASIC baseline against which WISE's fJ/MAC and TOPS/W claims are measured."},{"cited_title":"LeCun, L","cited_arxiv_id":null,"evidence_quote":"Provides the LeNet-300-100 architecture and the MNIST dataset used as the main classification benchmark."},{"cited_title":"Chu, Polyphase codes with good periodic correlation properties","cited_arxiv_id":null,"evidence_quote":"Defines the Zadoff-Chu sequence used in the activation function to spread waveform power across subcarriers."},{"cited_title":"Coaxial frequency mixer, 300–4300 MHz","cited_arxiv_id":null,"evidence_quote":"Is the passive double-balanced mixer hardware that performs the in-physics multiplication in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 1 pJ/sample ADC energy assumed in the energy-efficiency model."},{"cited_title":"Becker, J","cited_arxiv_id":null,"evidence_quote":"Supplies the AudioMNIST spoken-digit dataset used for the second classification task."}],"review_version":1}