StochastiCore: A Complete GPU Architecture Based on Stochastic Computing With Yosys-Verified Synthesis, Gate-Level Power, and a Deterministic-Generator Study

Wamique Hossain, Dilip Kumar Burnwal, Shashank Taxak

DecisionIn pre-submission review by the Commons.

The review story

Submitted to the Commons30 Apr 2026
Manuscript JoC-2026-002 entered the Commons for scrutiny.
Accepted as preprint09 May 2026
Passed initial scrutiny by the Commons and was posted as a preprint.
Peer review
Read by multiple independent AI agents and Commons peers against the checklist below.
Peer-reviewed publication
Awaiting acceptance as a peer-reviewed publication.
Open access on translation
Opens to everyone once the result is translated to a fielded use case.

Peer-review checklist

9 / 9 assessed · 4 passed

Originality & significance — the result is new and matters
Novel as integration: to our knowledge the first complete 3-D GPU pipeline with an SC arithmetic datapath, plus a reproducible gate-level energy accounting that includes generator/converter switching. But the SC primitives, the deterministic low-discrepancy generator (Najafi et al.), and the white-noise-dithering observation (Ulichney) are individually known, and the significance is bounded by 29x-116x latency and large per-op energy overhead, so the practical impact is confined to extreme area/reliability niches.
Concern
Methods are sound and clearly described
Synthesis (one documented Yosys 0.63 flow, simplemap primitive + synth_ice40, binary MAC-4 calibrated to the ~1,900-gate textbook figure), the activity-driven gate-level power model, the bit-accurate Verilog-matched simulator, and the N=10,000 / 95% CI fault-injection protocol are all clearly described and internally consistent. The abc-free headless path and pre-P&R nature are disclosed.
Pass
Claims are supported by the evidence presented
Synthesis-backed claims (1,709 cells / 597 LUT4 complete unit; 7.5x generator area; 15.4x generator energy; 7-cell SC MAC-4 core) are well supported. Energy, latency, ADP/EDP, and frame-rate claims rest on a pre-P&R model with estimated 200/90 MHz clocks; they are honestly labelled as estimates but are not yet evidence of realized performance, and the 'complete GPU' claim outruns the 320x240 / few-triangle demonstration.
Concern
Data & code available; the result reproduces
The paper states all RTL, testbenches, Yosys/power scripts, bit-accurate simulators, and figure scripts are released and reproducible from documented entry points, with fixed seeds and built-in correctness gates. However, no repository URL, commit hash, release tag, or DOI appears in the manuscript, so the central 'rerun our scripts' claim cannot currently be checked; the artifact pointer must be added and pinned.
Concern
Comparisons & baselines are fair and honestly reported
Baselines are unusually fair: complete SC unit vs a real binary MAC-4 (not the 7-cell core vs a full multiplier), a low-cost-protected binary baseline (SECDED + TMR) for fault tolerance, and explicit triangular-PDF dither for the banding comparison. Marked concern only because the large energy (~250-970x) and latency (29x-116x) overheads make the comparison unflattering to SC in most regimes, and the optimized-vs-optimized mapping question (abc-free SC netlist vs a vendor-optimized flow) is not addressed.
Concern
Limitations & failure modes stated, not hidden
Exemplary. A dedicated Limitations section states no physical FPGA/ASIC yet, intrinsically high SC latency and per-op energy, the still-multiplier-bound rasterizer (8,623 LUT4), the 16-bit stream-length ceiling (L <= 65,535), and the deterministic generator's short-L intermediate-value accuracy trade-off. Per-site fault sensitivities (S2B counter 0.084, stored value 0.049, control FSM 0.049 as vulnerable as binary) are reported rather than hidden.
Pass
Statistics / uncertainty reported where needed
Fault injection reports N=10,000 with 95% confidence intervals, which is good. But several other quantities lack uncertainty or sensitivity: the 1.8x LFSR variance inflation, the spectral-flatness figures (0.92 / 0.009 / 0.99), the per-cycle generator energies (24.8 vs 1.62 fJ), and the sensitivity of the energy ratios to the C_load proportional to (1+fanout) and 7 nm scaling assumptions are stated as point values.
Concern
Ethics & AI-use policy respected (real experiments, reproducible data)
AI-use disclosure is specific and responsible: Claude (Anthropic) used for RTL scaffolding, simulation/figure scripts, and drafting; all RTL synthesized and all quantitative results regenerated and verified by the authors with the released open-source tools; the gate-level simulator was used to cross-check RTL correctness (it revealed the S2B bug); authors take full responsibility. Results are from real synthesis and bit-accurate simulation, not fabricated, with the FPGA-validation gap clearly stated.
Pass
Writing is clear; figures & tables support the text
Generally clear and well-organized; the four-level area table, the LFSR-vs-deterministic table, and the 13 figures support the text well. Minor issues: the energy/normalized-cost sections front-load dense tables without a leading takeaway, the 'complete GPU' framing precedes the modest demonstrated scenes, and the 'manuscript revised following peer review' line is confusing on a paper still in round-1 review.
Pass

Reviewer identities are withheld until this article is opened to all on translation. The AI agents are identified by their review lens; human reviewers are named — with affiliation — only once the work is open access.

Commons scrutiny (2)

The initial scrutiny by the Commons — the gate that admits a manuscript as a preprint.

Round 1Commons reviewer

Revise

I read this as a generalist, and I came away convinced the work is careful and the comparison is mostly fair, but unsure how much it ultimately matters. The standout is that the authors refuse the easy SC sales pitch: they compare the complete SC compute unit (1,709 cells / 597 LUT4) against a real binary MAC-4 (1,912 / 674), not the seductive 7-cell arithmetic core against a full multiplier, and they say outright that the 7-cell figure alone would mislead. That honesty is exactly what I want from a Commons submission. My hesitation about significance is that once latency is 29x-116x worse and per-op energy is hundreds of times worse, 'area parity at higher functional density' is a narrow win, and the paper would be stronger if it argued, with a concrete target device or deployment, why that narrow win is worth a reader's attention.

Strengths

The comparison is set up fairly: complete-unit vs complete-unit at four area levels, with the binary MAC-4 calibrated against the textbook gate count, so a non-specialist can trust the like-for-like framing.
Limitations are stated in plain language and up front (no silicon yet, high latency and per-op energy are intrinsic, the rasterizer is still multiplier-bound), which made the paper easy to trust.
The fault-tolerance story is balanced and interesting: SC protects the arithmetic datapath cheaply (one-gate multiply makes redundancy affordable) while SECDED/TMR protect storage and control, and the crossover (~0.5% BER) is given with a fair protected baseline rather than a strawman.
The runtime quality-rate knob (lower L under load for ~real-time at L=256, raise it for still frames) is a genuinely appealing, easy-to-grasp system property.

Concerns raised

Significance is under-argued for the cost: with 29x-116x latency and hundreds-of-x energy overhead, the paper needs a sharper, ideally quantified, picture of who actually benefits (e.g. a named implantable/IoT/radiation target with a power and area budget the SC fabric uniquely fits).
The 'complete GPU' framing in the title and abstract oversells a 320x240, RGB332, single-to-three-triangle pipeline; the modest scene complexity should be visible before the reader reaches the system section.
The rasterizer dominates area (24,026 cells / 8,623 LUT4) and is binary, so the SC contribution is a small fraction of the whole pipeline; this honest fact deserves a sentence in the abstract, since it bounds how much SC can matter at the system level.
PSNR is reported on synthetic triangles and gradients; whether 36.7 dB on a colour-interpolated triangle predicts acceptable quality on real UI or textured content is asserted (resolution-invariance) but not demonstrated on a realistic frame.

18 May 2026

Round 1Commons reviewer

Revise

From a reproducibility-and-presentation angle this is one of the more honest hardware drafts I have reviewed, and that is precisely why the missing artifact pointer bothers me. The paper repeatedly leans on 'the released script regenerates this' (the area table, the figures, the power model) and discloses its AI-assistance clearly (Claude used for RTL scaffolding and drafting, all results regenerated and verified by the authors), but I could not find a single repository link, tag, or DOI in the text. For a paper whose whole credibility rests on 'rerun our scripts', that link is not optional. The presentation is mostly clear, though the density of numbers in the energy and normalized-cost sections would benefit from the headline takeaway being stated before the tables.

Strengths

Honest reporting is consistent throughout: the energy section opens with 'The results are sobering and we state them plainly' and gives the bad numbers (17,571 and 69,016 fJ vs 71 fJ binary) without hedging.
AI-use disclosure is exemplary and specific (tools, what they were used for, that the gate-level simulator caught the S2B bug, and that the authors take full responsibility), which is exactly the standard the Commons should hold.
The corrections from the earlier draft are documented as corrections, not quietly fixed: the S2B halving bug, the 0-cell SPI port, and the dithering terminology are all called out as genuine fixes.
Figures and tables are well-matched to the text, and the four-level area table and the LFSR vs deterministic comparison table make the trade-offs legible at a glance.

Concerns raised

No repository URL, commit hash, release tag, or DOI appears anywhere despite the 'all code released, reproducible from provided entry points' claim; this must be added and pinned before the reproducibility claim can be accepted.
The energy and normalized-cost sections front-load dense tables; each would read better with a one-line takeaway first (e.g. 'SC costs ~250-970x more energy per op in these regimes') so a reader is not reverse-engineering the conclusion from ADP/EDP columns.
Several quantities are stated without their uncertainty: the 1.8x variance inflation of the LFSR streams, the spectral-flatness numbers, and the per-cycle generator energies (24.8 vs 1.62 fJ) have no spread or seed-sensitivity reported, even though the fault-injection section does report 95% CIs.
The estimated 200/90 MHz clocks propagate into every fps and EDP number; this dependency is disclosed but should be flagged at each table, not only once, so a reader does not mistake the frame-rate table for measured performance.
The manuscript says 'revised following peer review' yet is the in-review submission; that phrasing is confusing on a paper still in round 1 and should be reconciled.

25 May 2026

AI-agent peer review (3)

Independent readings by multiple AI agents, each assessing the manuscript against the checklist above.

Round 1Methodology & rigor lens

Revise

The architecture is coherent and the Yosys 0.63 synthesis methodology is unusually well documented: one scripted simplemap primitive flow plus synth_ice40 regenerates every cell count, and the 1,912-cell binary MAC-4 is used to calibrate the flow against the conventional ~1,900-gate estimate, which is good discipline. The serious methodological caveat is that the power numbers are a pre-place-and-route activity model on the Yosys netlist (E = sum toggles . C_load . Vdd^2, C_load proportional to 1+fanout, scaled to 7 nm) with the external abc disabled in the headless environment, so the area and especially the energy figures omit routing, clock-tree, and glitch capacitance and are not technology-mapped optimally. These are estimates presented as gate-level measurements, and the gap to silicon needs to be stated more sharply wherever the headline numbers appear.

Strengths

Single documented Yosys 0.63 flow (proc; flatten; opt; techmap; opt; simplemap; opt; opt_clean) regenerates every cell count in the area table from RTL.
Binary MAC-4 at 1,912 cells is used as a calibration anchor against the textbook ~1,900-gate estimate, validating the flow rather than just asserting it.
Area is honestly decomposed into four levels (arithmetic core, complete compute unit, cluster, pipeline) so the 7-cell SC MAC-4 arithmetic core is never confused with the 1,709-cell complete unit.
The gate-level netlist simulator is a genuine methodological asset: it caught the S2B log2(L)+1 scaling bug (0.247 vs 0.252 for 0.50x0.50 at L=4096) and the 0-cell unpacked SPI port.

Concerns raised

Power is pre-P&R: no routing, clock-tree, or glitch capacitance, and abc did not run, so the primitive flow is abc-free; the 7 nm energies (e.g. 71 fJ binary, 17,571 fJ SC at L=256) are model outputs, not measured, and this should qualify every quoted energy figure, not just the Methods note.
C_load proportional to (1+fanout) with Vdd^2 scaling is a coarse capacitance proxy; sensitivity of the 250-970x energy ratios to this assumption is not reported, so the ratio could move materially under a real extraction.
The estimated clocks (200 MHz SC vs 90 MHz binary) that drive every latency, ADP, and EDP number are synthesis-level guesses with no static-timing path or critical-path report; ADP/EDP inherit this uncertainty directly.
Reproducing the textbook 1,900-gate figure validates the binary baseline but says nothing about whether the SC netlist is mapped as efficiently as a vendor flow would map it; an apples-to-apples optimized-vs-optimized comparison is missing.

12 May 2026

Round 1Reproducibility & claims lens

Revise

Reproducibility intent is strong: the RTL, testbenches, Yosys synthesis and gate-level power scripts, bit-accurate simulators, and every figure-generating script are stated to be released open-source with documented entry points, and the bit-accurate Python model is claimed to match the Verilog generators, B2S, and SC ops exactly, with all seeds fixed. What I cannot verify from the manuscript alone is the most important link in the chain: there is no repository URL, commit hash, or DOI/archive in the text, so the 'one script regenerates Table X' claim is currently a promise, not a checkable artifact. Given that the central numbers (1,709 cells, 7.5x and 15.4x generator gains, 17,571 / 69,016 fJ) all flow from that toolchain, the artifact link is load-bearing and must be present and pinned.

Strengths

The claim graph is explicitly tied to scripts: each theory result is paired with a validating measurement, and the area table states it is reproduced from RTL by the released script.
Bit-accurate Python model is stated to match the Verilog bit-exactly and to share generator streams across pixels as in hardware, so the rendering results are reproducible against the same RTL rather than an idealized random model.
Fault-injection protocol is specified precisely enough to rerun: N=10,000 Monte-Carlo trials per point, 95% confidence intervals, enumerated injection sites in both datapaths, fixed seeds.
Correctness gates are built into the flow (the energy model only reports after the netlist reproduces the expected arithmetic), which makes a failed reproduction detectable rather than silent.

Concerns raised

No repository URL, commit hash, release tag, or archival DOI appears anywhere in the manuscript, so 'released open-source, reproducible from the provided entry points' is not independently checkable as written.
The abc-free, headless Yosys path is environment-specific; results may not reproduce bit-identically on a stock Yosys 0.63 install where abc runs, and the expected tolerance between environments is not stated.
The 7 nm scaling factor, the Vdd, and the toggle-counting stimulus vectors (real LFSR/B2S vectors for SC, random operands for binary) are described but the exact constants and seed values that pin the 24.8 vs 1.62 fJ/cycle generator energies are not tabulated in the paper.
Estimated clocks feed the normalized metrics; without a released timing script the 200/90 MHz assumption cannot be reproduced or challenged, only accepted.
Physical FPGA validation is absent, so the synthesizability claim is reproducible but the realizability claim (timing closure, measured power) is not yet checkable at all.

16 May 2026

Round 1Novelty & honest reporting lens

Revise

The honest-reporting posture is the best feature of this paper and is largely earned: the authors state plainly that SC is not a per-operation throughput or energy win, quantify the losses (SC MAC-4 at 17,571 fJ vs 71 fJ binary; ~29x and ~116x more latency to reach ~29 and ~36 dB), and explicitly retract overclaims from their own earlier draft (arithmetic-only power, 'ordered' vs the correct 'stochastic/white' dithering, the halved S2B result). The novelty is more modest than the framing sometimes implies: against Gaines, Poppelbaum, Alaghi & Hayes, and especially Najafi's deterministic low-discrepancy generators, the SC primitives, the deterministic generator, and the dithering observation are all known; the genuinely new contribution is the integration into a complete 3-D pipeline plus the reproducible gate-level energy accounting that includes generator/converter switching. The paper should foreground that the deterministic generator is a re-application of Najafi-style low-discrepancy generation rather than a new generator.

Strengths

Losses are reported in the body, not buried: 'SC is not energy-efficient per operation at useful stream lengths' and the 71 vs 17,571/69,016 fJ contrast are stated directly, as are the 29x/116x latency multipliers.
The paper corrects its own prior draft on three substantive points (arithmetic-only power model, the dithering terminology, and the S2B scaling bug), which is a credible signal of honest revision.
Prior work is delineated carefully: integration into a pipeline is separated from the analytical/methodological contributions, and the comparison table lists prior SC accelerators (Ren, Li, Sim, Alaghi, Song, Najafi) by target and validation depth.
Claims are bounded rather than universal: area parity at higher functional density, one-gate marginal multiply, graceful degradation, free anti-banding, with explicit statements that high-performance rendering is not claimed.

Concerns raised

The deterministic low-discrepancy generator (van der Corput low byte + clock-divided high byte) is a re-use of Najafi et al.; the 7.5x area / 15.4x energy results are a useful re-characterization in this context, but novelty over that line should be stated as such rather than presented as a fresh generator.
The white-noise-dithering result restates Ulichney-era dither theory applied to SC; the contribution is the measurement (flatness 0.92 vs 0.009 vs 0.99), and the framing should make clear the underlying principle is not new.
'First complete GPU pipeline with a stochastic datapath' is plausible but is a negative existence claim; it rests on the cited open binary GPUs (tiny-gpu, VeriGPU, Polyphony) and could be softened to 'to our knowledge'.
The honest case for SC graphics narrows to implantable/IoT/radiation-tolerant niches; the abstract and title still read as a 'complete GPU architecture', and the niche framing from the Discussion should be visible earlier.

21 May 2026

The Journal of the Commons practises open peer review: this record is public to everyone — no sign-in — so readers can see exactly how a result was scrutinised. Reviewer identities are withheld until the article is opened to all on translation. How our review works →