{"id":"3b5f48e7-e78b-4410-9189-40904ca03389","arxiv_id":"2504.17883","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PowerSensor3 is an open, modular power measurement instrument with 20 kHz sampling, isolated Hall current sensors, and sub-EUR 100 component cost, validated on GPUs, a Jetson SoC, and an SSD.","lead":"PowerSensor3 is an open-source hardware and software toolkit that measures the power draw of GPUs, SSDs, and other PCIe devices at up to 20,000 samples per second. A smart generalist might read it because cheap, high-resolution external power sensors could make energy-efficiency optimization of AI and HPC systems practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Measured 20 kHz peak-to-peak error exceeds Table I's ±4.2 W 'theoretical worst case' bound, making the headline accuracy claim internally inconsistent as stated.","rationale":"The reader correctly identifies that the quantitative accuracy claim rests on datasheet noise figures and that the paper's own Table II measurements appear to contradict Table I. My independent reading confirms this is the most load-bearing soft spot: the central claim is 'fast and accurate', and 'accurate' is quantified by a worst-case bound that the paper's measured data at 20 kHz exceeds. The contradiction is concrete, located in specific tables, and testable. The rest of the paper is strong: open hardware/software artifacts, reproducible results, a clear calibration procedure, and use cases (GPU transients, SSD power, kernel tuning) that support the qualitative value of the tool even if the numerical bound is restated. Therefore the concern does not invalidate the paper, but it does require a corrected or clarified accuracy statement, keeping the reader's CONDITIONAL verdict. No change to the verdict is needed because the reader already flagged this issue.","tokens_in":21589,"tokens_out":3869,"duration_ms":36751,"concrete_test":"Recompute Table I from first principles and compare to Table II: with 115 mArms at 12 V, the noise contribution is 1.38 Wrms; a 3-sigma peak-to-peak estimate is about 4.14 Wpp, while the observed 20 kHz max-min at 0.5 A is 6.38 Wpp. Then rerun the Table II experiment at the stated 8 A load and at 0.5 A/1 A using a calibrated shunt/current probe as reference, and report both RMS error and peak-to-peak error at each sampling rate. If the measured p-p exceeds the quoted Table I bound, the bound should be relabeled (e.g., RMS or 3-sigma) and the headline accuracy restated accordingly. This single check settles whether the accuracy claim holds as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A derives a 'theoretical worst case' accuracy for the 12 V/10 A module of ±4.2 W (Table I) from the MLX91221 Hall sensor's 115 mArms noise and the ACPL-C87B voltage noise. This bound is load-bearing for the paper's 'accurate' central claim. Section IV-A's own 20 kHz measurements (Table II) show a peak-to-peak error of 6.38 W at a 0.5 A load and 7.685 W at 1 A, both larger than ±4.2 W interpreted as a half-range or as a 4.2 W peak-to-peak figure. The table caption states an '8 A load' while the table lists 0.5 A and 1 A loads, so the exact measurement condition is unclear; if the caption is the typo, the 0.5 A/1 A rows still contradict Table I unless 'theoretical worst case' means something narrower (e.g., 3-sigma RMS) than the measured peak-to-peak range. The paper does not reconcile this. Since the headline claim quantifies worst-case accuracy, an unaddressed discrepancy between stated bound and measured 20 kHz noise is a real correctness risk. The use-case demonstration that 20 kHz reveals transients NVML misses (Section V-A, Fig. 7a) is independent and credible; the concern is specifically the quantitative accuracy bound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents PowerSensor3, an open-source hardware/firmware/software toolkit for external power measurement of PCIe devices, SoC boards, and peripheral cards. The hardware uses a modular baseboard with up to four sensor modules, a Hall-based current sensor, an optically isolated voltage sensor, and an STM32 microcontroller sampling at up to 20 kHz. The authors characterize sensor accuracy with a current sweep against Fluke reference instruments, a 50-hour stability test, and step-response measurements. They then demonstrate the tool in three use cases: NVIDIA and AMD GPU power profiling, NVIDIA Jetson AGX Orin measurement, and SSD power analysis. The central claims are that the 12 V/10 A module achieves a theoretical worst-case power accuracy of ±4.2 W, that 20 kHz sampling reveals transients missed by NVML, and that integration with Kernel Tuner reduces auto-tuning time by 3.25x.","tokens_in":21892,"tokens_out":9835,"duration_ms":94436,"significance":"If the accuracy and sampling claims hold, PowerSensor3 is a valuable open tool for energy-efficiency research, filling a gap left by vendor sensors with ~10 Hz update rates and by expensive or non-open commercial instruments. The paper's strengths include the open-hardware and open-software releases with archived DOIs, the use of external reference instruments for accuracy evaluation, the 50-hour stability check, and real application demonstrations on GPUs, an SoC, and an SSD. The 20 kHz capability is demonstrated by visible power transients that NVML misses, which is a concrete and falsifiable contribution. The main weakness is that the accuracy numbers in Section III-A and Section IV-A are stated without a precise statistical definition, which currently makes the headline 'accurate' claim difficult to verify.","major_comments":[{"comment":"Table II's caption and column semantics must be fixed before the accuracy claim can be evaluated. The caption says 'error values' and '8 A load', but the rows are 0.5 A and 1 A loads, and the min/max entries are centered on 6 W and 12 W, respectively, indicating that they are measured power readings rather than signed errors. If that is the case, the 20 kHz peak-to-peak ranges of 6.38 W (0.5 A) and 7.685 W (1 A) correspond to worst-case absolute deviations of roughly 3.22 W and 4.21 W from the expected powers, which would be within the Table I '±4.2 W' bound; the paper should state this explicitly. If the entries really are errors, the large positive values imply a bias that is not discussed elsewhere. Please rewrite the caption, define min/max/p-p/std, give the expected power at each load, and correct the '8 A load' statement.","section":"IV-A, Table II; III-A, Table I"},{"comment":"The 'theoretical worst-case accuracy' in Table I is not derived in a checkable way. The text reports a Hall-sensor noise of 115 mArms and calls 4.1 W the resulting 'peak-to-peak error', yet Table I lists current error ±0.35 A and power ±4.2 W. Please state the assumed conversion from RMS to peak-to-peak (or define the confidence level), the operating point (nominal voltage and current) used in E_p = sqrt((U*E_i)^2 + (I*E_u)^2 + (E_i*E_u)^2), and how quantization, calibration offset, and voltage-divider noise enter the ±28.6 mV and ±0.35 A entries. Without this definition, the headline accuracy cannot be checked against the measurements in Section IV.","section":"III-A, Table I"}],"minor_comments":[{"comment":"Report the numerical drift in the minimum and maximum power values over the 50-hour test, not only the ±0.09 W fluctuation of the averages, to support the 'no recalibration needed' claim.","section":"IV-B"},{"comment":"Quantify the step response with rise time and settling time rather than only showing the waveform, so readers can compare the dynamic performance with other instruments.","section":"IV-C"},{"comment":"The 3.25x tuning-time comparison depends on an estimated 7394 s for the onboard-sensor method; state the estimation assumptions, such as the number of repeated kernel executions and the assumed run duration per configuration.","section":"V-A2"},{"comment":"The axis labels in Figure 4 are garbled in the current PDF; ensure the figure is legible and that the axis label clearly states whether the plotted quantity is power error or measured power.","section":"Fig. 4"},{"comment":"The claim that the current sensors are 'hardly sensitive' to external magnetic fields is not experimentally demonstrated; a short comparative test with a nearby current-carrying conductor would support this design claim.","section":"III-A"}],"recommendation":"major_revision","confidential_remarks":"This is a solid systems/tool paper that fits the venue. The main blocker is the precision of the accuracy definitions: the apparent contradiction between the ±4.2 W bound and the Table II peak-to-peak values is likely resolvable by clarifying that Table II's min/max are measured power readings, but this must be made explicit and the caption corrected. Once that is done, I expect the paper to be acceptable; no novelty or scope concerns beyond the authors' own PowerSensor2 prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the hardware is real and the artifacts are genuinely open: hardware under CERN-OHL-P v2, firmware and software under Apache-2.0, plus evaluation data on Zenodo. That is the right way to ship a measurement tool. Second, the headline accuracy claim has a soft spot that a referee will catch: Table I gives a theoretical worst-case power accuracy of ±4.2 W for the 12 V/10 A module, but Table II's own 20 kHz measurements show peak-to-peak errors of 6.38 W at 0.5 A and 7.69 W at 1 A. The caption even says \"8 A load\" while the rows are 0.5 A and 1 A. That inconsistency is not fatal to the tool's usefulness, but it needs a clear reconciliation or a revised bound.\n\nWhat is actually new: PowerSensor3 upgrades the authors' PowerSensor2 from 2.8 kHz to 20 kHz sampling, adds modular sensor boards, isolated Hall current sensors, simplified calibration, and a sub-100 Euro bill of materials. The combination of 20 kHz, multi-rail, isolated measurement in an open design is not in the prior literature. The paper also does solid evaluation work: a current sweep against Fluke references, a 50-hour stability test, step-response measurements, and three realistic case studies. The GPU transient plots showing NVML missing power dips are convincing, and the Kernel Tuner integration producing a 3.25x wall-clock speedup is a concrete, reproducible benefit.\n\nThe soft spots are proportionate. The Table I vs Table II discrepancy is the main one; it may be that \"theoretical worst case\" is a 3-sigma RMS bound while the table reports the sample peak-to-peak, but the paper never says that. A one-sentence clarification plus a corrected caption would address most of it. The accuracy sweep is also only shown in aggregate for a 12 V sensor; some high-current data points would help, since the error formula says voltage noise dominates there. The self-citations to PowerSensor2, PMT, and Kernel Tuner are fine because the prior work is real and the new contribution is the device itself. The case studies are applications, not the core contribution, so I would not judge the paper on their breadth.\n\nWho should read this: anyone doing GPU, SSD, or SoC energy-efficiency research or auto-tuning who needs sub-millisecond external power measurement. It deserves a serious referee, and a revised version with a tightened accuracy argument would be worth publishing. My verdict: send it to review, with a request to fix the caption and reconcile the accuracy numbers.","headline":"A genuinely useful, openly released 20 kHz power sensor for GPU/SSD energy work, with a real but fixable inconsistency between the theoretical accuracy bound and the measured low-load noise.","tokens_in":22457,"tokens_out":1223,"would_cite":true,"duration_ms":14868,"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":"PowerSensor3 is an open-hardware measurement tool that samples power at 20 kHz and reveals GPU transients that built-in sensors updating at about 10 Hz miss.","keywords":["power measurement","open hardware","GPU power profiling","20 kHz sampling","PCIe power","energy efficiency","current sensor","auto-tuning"],"falsifier":"Measure a constant 0.5 A load on a 12 V / 10 A module with a calibrated laboratory supply and electronic load, and record at 20 kHz for several seconds; if the peak-to-peak spread of power readings consistently exceeds the stated worst-case accuracy of roughly ±4.2 W, as the paper's Table II already suggests, then the accuracy claim as stated is falsified.","tokens_in":21416,"feed_emoji":"⚡","tokens_out":7399,"duration_ms":64922,"temperature":0.7,"pith_summary":"PowerSensor3 is an open-hardware, open-software measurement tool for the power consumption of PCIe devices and embedded boards. It samples voltage and current at 20 kHz through modular sensor boards and derives power as their product, with a theoretical worst-case accuracy of ±4.2 W for its 12 V / 10 A module. The paper argues that this speed matters: on one recent GPU, PowerSensor3 showed clear power dips between the phases of a single kernel, while the GPU's built-in sensor, updating near 10 Hz, missed them; on a GPU from another vendor, the external measurements closely tracked the built-in sensor. The paper also shows that powering an auto-tuner with this measurement backend cut the time needed to tune a beamforming application for energy efficiency by 3.25 times. The designs, firmware, and host libraries are released openly, putting sub-millisecond power measurement within reach of researchers without commercial instrumentation.","feed_headline":"20 kHz power meter catches GPU transients built-in sensors miss","feed_subtitle":"An open-hardware design under $100 in parts measures PCIe devices and shows the power dips a 10 Hz GPU sensor cannot catch.","key_machinery":"The central machinery is the modular sensor chain: a baseboard with up to four interchangeable sensor boards, each carrying a differential Hall current sensor and an optically isolated voltage amplifier, feeding a 24 MHz ADC that averages six samples per channel to emit 20 kSamples/s over USB, with a remote-sense connector moving the voltage reference point to the load. The accuracy argument rests on the error-propagation identity $E_p = \\sqrt{(U E_i)^2 + (I E_u)^2 + (E_i E_u)^2}$ derived from $P = (U+E_u)(I+E_i)$, where $E_i$ is dominated by the Hall sensor's 115 mArms datasheet noise and $E_u$ by quantization plus amplifier noise through the voltage divider.","core_discovery":"The central claim is that a cheap, modular, open design can deliver sub-millisecond power measurements accurate enough to make GPU power visible at kernel granularity. Each sensor board pairs a differential Hall current sensor with an optically isolated voltage sensor, and the microcontroller streams 20 kSamples/s to the host over USB. The paper characterizes the error budget with $P = (U+E_u)(I+E_i)$, so the power error is $E_p = \\sqrt{(U E_i)^2 + (I E_u)^2 + (E_i E_u)^2}$; at low loads current-sensor noise dominates, and at high currents voltage-sensor noise dominates. In the GPU case study the measured signal shows transient phases and dips that the vendor's roughly 10 Hz readings do not, and in the tuning case study the 3.25x speedup comes from measuring each kernel variant once instead of running it for seconds to gather enough samples. For devices without built-in power instrumentation, such as PCIe SSDs and SoC boards, the tool supplies measurements where none existed.","pith_inferences":["A natural next step the paper does not take is to use the 20 kHz channel to build per-operation or per-kernel power models on GPUs, since phase boundaries visible in the traces can be aligned with kernel launch and memory events.","The apparent gap between the stated ±4.2 W bound and the measured 6.38 W peak-to-peak spread at a 0.5 A load suggests that adopting a statistical definition of accuracy, such as a confidence interval on the distribution of readings, would make the headline claim testable; until then, the bound is best read as a design target.","Because the sensor boards are interchangeable, synchronized measurements on multiple rails or multiple baseboards could give whole-node power attribution per component, which the present paper only hints at with its four-sensor baseboard.","For storage, 20 kHz sampling could expose per-command power spikes in SSDs that one-second averages hide; the paper mentions future sub-millisecond SSD work, so this is an open direction."],"forward_implications":["GPU energy accounting can be done at kernel or even sub-kernel granularity without artificially lengthening workloads, because the 20 kHz stream resolves transients that roughly 10 Hz built-in sensors average away.","Auto-tuners that optimize for energy can benchmark each code variant directly and briefly, cutting tuning time; the paper reports a 3.25x reduction for a beamforming application.","PCIe devices without power instrumentation, such as NVMe SSDs, NICs, FPGAs, and domain-specific accelerators, can be monitored externally at sub-millisecond resolution, making SSD power behavior visible independently of bandwidth.","The open, modular design and one-time calibration mean a research group can reproduce the hardware, swap sensor boards for different power ranges, and integrate the measurements into its own software stack."],"supporting_citations":[{"why":"Describes the previous device whose sampling rate and design PowerSensor3 builds on and improves.","marker":"[27]"},{"why":"Documents the low update rate and inaccuracies of a GPU's built-in power sensor, establishing the comparison baseline for PowerSensor3.","marker":"[38]"},{"why":"Datasheet supplying the Hall current sensor's noise specification used in the theoretical error budget.","marker":"[48]"},{"why":"Datasheet supplying the isolated voltage sensor's performance characteristics used in the error budget.","marker":"[49]"},{"why":"Software used to compare PowerSensor3 readings against vendor power APIs on GPUs.","marker":"[51]"},{"why":"Auto-tuner into which PowerSensor3 was integrated to measure kernel energy directly and demonstrate the 3.25x tuning speedup.","marker":"[54]"},{"why":"Prior study of SSD power that PowerSensor3's SSD case study extends with standardized external sensing.","marker":"[58]"}],"fun_headline_variants":["Sub-$100 open hardware power meter catches GPU dips","20 kHz open-source sensor reveals GPU power transients","Open-source meter samples 20 kHz, sees GPU power dips","20 kHz open hardware uncovers GPU transients built-ins miss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy figures assume the Hall current sensor's datasheet noise of 115 mArms and the isolated voltage sensor's datasheet performance are realized in the assembled, cabled module; if those specifications do not hold in practice, the quoted ±4.2 W worst-case accuracy does not follow, and the paper's own 20 kHz measurements at a 0.5 A load (6.38 W peak-to-peak) already look inconsistent with that bound.","fun_headline_variants_meta":{"raw":{"variants":["Sub-$100 open hardware power meter catches GPU dips","20 kHz open-source sensor reveals GPU power transients","Open-source meter samples 20 kHz, sees GPU power dips","20 kHz open hardware uncovers GPU transients built-ins miss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001735,"raw_usage":{"total_tokens":6889,"prompt_tokens":1008,"completion_tokens":5881,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":5814}},"tokens_in":624,"tokens_out":5881,"duration_ms":33951,"temperature":1.0,"reasoning_tokens":5814,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:30:11.315625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure a constant 0.5 A load on a 12 V / 10 A module with a calibrated laboratory supply and electronic load, and record at 20 kHz for several seconds; if the peak-to-peak spread of power readings consistently exceeds the stated worst-case accuracy of roughly ±4.2 W, as the paper's Table II already suggests, then the accuracy claim as stated is falsified.","supporting_citations":[{"cited_title":"PowerSensor 2: A Fast Power Mea- surement Tool,","cited_arxiv_id":null,"evidence_quote":"Describes the previous device whose sampling rate and design PowerSensor3 builds on and improves."},{"cited_title":"Accurate and Convenient Energy Measurements for GPUs: A Detailed Study of NVIDIA GPU’s Built- In Power Sensor,","cited_arxiv_id":null,"evidence_quote":"Documents the low update rate and inaccuracies of a GPU's built-in power sensor, establishing the comparison baseline for PowerSensor3."},{"cited_title":"Datasheet MLX91221 Integrated Current Sensor IC,","cited_arxiv_id":null,"evidence_quote":"Datasheet supplying the Hall current sensor's noise specification used in the theoretical error budget."},{"cited_title":"Datasheet ACPL-C87B Precision Optically Isolated V oltage Sensor,","cited_arxiv_id":null,"evidence_quote":"Datasheet supplying the isolated voltage sensor's performance characteristics used in the error budget."},{"cited_title":"PMT: Power Measurement Toolkit,","cited_arxiv_id":null,"evidence_quote":"Software used to compare PowerSensor3 readings against vendor power APIs on GPUs."},{"cited_title":"Kernel Tuner: A search-optimizing GPU code auto- tuner,","cited_arxiv_id":null,"evidence_quote":"Auto-tuner into which PowerSensor3 was integrated to measure kernel energy directly and demonstrate the 3.25x tuning speedup."},{"cited_title":"Can Storage Devices be Power Adaptive?","cited_arxiv_id":null,"evidence_quote":"Prior study of SSD power that PowerSensor3's SSD case study extends with standardized external sensing."}],"review_version":1}