arxiv: 2604.03432 · v1 · submitted 2026-04-03 · 💻 cs.NE · cs.AR

Recognition: 2 theorem links

· Lean Theorem

YANA: Bridging the Neuromorphic Simulation-to-Hardware Gap

Brian Pachideh, Carmen Weigelt, Jann Krausse, Juergen Becker, Klaus Knobloch, Moritz Neher, Sven Nitzsche, Victor Pazmino Betancourt

Authors on Pith no claims yet

Pith reviewed 2026-05-13 18:08 UTC · model grok-4.3

classification 💻 cs.NE cs.AR

keywords spiking neural networksFPGA acceleratorneuromorphic computingevent-driven pipelinesparsity exploitationSNN hardwareopen-source frameworkneuromorphic simulation

0 comments

The pith

YANA is an FPGA-based digital accelerator that uses a five-stage event-driven pipeline to exploit sparsity in spiking neural networks and bridge simulation to hardware.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents YANA as an open-source FPGA framework to close the gap between SNN simulations and actual neuromorphic hardware. It shows that a carefully designed event-driven pipeline can process inputs at one event per cycle while scaling inference time almost linearly as data becomes sparser in time or space. This approach matters because it lets researchers test and optimize spiking networks on readily available hardware instead of waiting for scarce neuromorphic chips. The design supports any network topology through direct connections and keeps resource use low enough for small FPGA boards. Releasing the code aims to speed up progress in low-power, event-based computing.

Core claim

YANA implements a five-stage, event-driven processing pipeline on FPGA that fully exploits temporal and spatial sparsity while supporting arbitrary SNN topologies through point-to-point neuron connections. An input preprocessing scheme ensures steady one-event-per-cycle throughput without buffer overflow, and lookup tables handle leak calculations efficiently. On the Spiking Heidelberg Digits dataset, inference time scales near-linearly with both spatial and temporal sparsity levels. The core uses 740 LUTs, 918 registers, 7 BRAMs and 24 URAMs on the AMD Kria KR260, supporting up to 2^17 synapses and 2^10 neurons, and the full framework is released open-source with NIR integration.

What carries the argument

five-stage event-driven processing pipeline with input preprocessing for one-event-per-cycle throughput and lookup tables for neuron leak calculations

If this is right

Inference time scales near-linearly with spatial and temporal sparsity levels on spiking datasets.
Arbitrary SNN topologies are supported via point-to-point connections without pipeline stalls.
Resource requirements stay low, fitting on accessible platforms like the AMD Kria KR260 with capacity for large synapse counts.
The open-source release enables integrated training, optimization and deployment workflows for neuromorphic applications.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Researchers could prototype and debug SNN algorithms on FPGAs before targeting specialized neuromorphic chips, shortening development cycles.
The sparsity scaling observed may generalize to other event-driven sensors such as vision or audio streams in real-time systems.
Standardizing through NIR integration could help different neuromorphic software tools interoperate more easily.
Multiple YANA cores might run in parallel on one FPGA to handle bigger networks or higher throughput.

Load-bearing premise

The five-stage event-driven pipeline will maintain one-event-per-cycle throughput and avoid buffer issues across arbitrary real-world SNN topologies and input rates beyond the tested dataset.

What would settle it

A test showing buffer overflows, event drops, or significantly sub-linear scaling when YANA processes a high-rate or complex SNN topology outside the Spiking Heidelberg Digits dataset.

Figures

Figures reproduced from arXiv: 2604.03432 by Brian Pachideh, Carmen Weigelt, Jann Krausse, Juergen Becker, Klaus Knobloch, Moritz Neher, Sven Nitzsche, Victor Pazmino Betancourt.

**Figure 2.** Figure 2: YANA architecture integration complete with AXI4 stream buffers, a control unit and cores for input, hidden and output computations. synchronized timestep, decodes and parses incoming commands and maintains an internal state machine to manage the control flow. To interface with the accelerator’s CU, we instantiate AXI-enabled buffers for the input data, control commands and result output. These are accessi… view at source ↗

**Figure 3.** Figure 3: Overview of the YANA software framework 3 YANA Software Framework When developing any hardware architecture, the accompanying software is the key to making it widely usable without large overheads stemming from custom deployment solutions. Especially in the DL community, there are established tools and workflows that many practitioners commonly use to develop their neural networks. Therefore, our goal for … view at source ↗

**Figure 4.** Figure 4: Scaling of inference time of SNNs with different spatial and temporal sparsity levels after deployment on YANA. In the case of sweeping Sspat, Stemp cannot be fixed and has to be given in a small range, since the pruning level influences hidden layer sparsity, ultimately changing the total Stemp. 4.2 Results [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

read the original abstract

Spiking Neural Networks (SNNs) promise significant advantages over conventional Artificial Neural Networks (ANNs) for applications requiring real-time processing of temporally sparse data streams under strict power constraints -- a concept known as the Neuromorphic Advantage. However, the limited availability of neuromorphic hardware creates a substantial simulation-to-hardware gap that impedes algorithmic innovation, hardware-software co-design, and the development of mature open-source ecosystems. To address this challenge, we introduce Yet Another Neuromorphic Accelerator (YANA), an FPGA-based digital SNN accelerator designed to bridge this gap by providing an accessible hardware and software framework for neuromorphic computing. YANA implements a five-stage, event-driven processing pipeline that fully exploits temporal and spatial sparsity while supporting arbitrary SNN topologies through point-to-point neuron connections. The architecture features an input preprocessing scheme that maintains steady event processing at one event per cycle without buffer overflow risks, and implements hardware-efficient event-driven neuron updates using lookup tables for leak calculations. We demonstrate YANA's sparsity exploitation capabilities through experiments on the Spiking Heidelberg Digits dataset, showing near-linear scaling of inference time with both spatial and temporal sparsity levels. Deployed on the accessible AMD Kria KR260 platform, a single YANA core utilizes 740 LUTs, 918 registers, 7 BRAMS and 24 URAMs, supporting up to $2^{17}$ synapses and $2^{10}$ neurons. We release the YANA framework as an open-source project, providing an end-to-end solution for training, optimizing, and deploying SNNs that integrates with existing neuromorphic computing tools through the Neuromorphic Intermediate Representation (NIR).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

YANA gives a practical open-source FPGA SNN accelerator with real measured sparsity scaling on SHD, but the one-event-per-cycle claim for arbitrary topologies is only lightly tested.

read the letter

YANA is a straightforward FPGA implementation of an event-driven SNN accelerator built on the AMD Kria KR260. The authors describe a five-stage pipeline, LUT-based leak updates, and an input preprocessing step meant to keep processing at one event per cycle. They report concrete resource numbers (740 LUTs, 918 registers, small BRAM/URAM counts) and show near-linear scaling of inference time with spatial and temporal sparsity on the Spiking Heidelberg Digits dataset. The open-source release and NIR integration are useful additions for anyone who wants to move SNN work from simulation onto actual hardware without waiting for scarce chips.

Referee Report

1 major / 2 minor

Summary. The paper introduces Yet Another Neuromorphic Accelerator (YANA), an FPGA-based digital SNN accelerator featuring a five-stage event-driven pipeline that exploits temporal and spatial sparsity while supporting arbitrary topologies via point-to-point connections. It includes input preprocessing to maintain one-event-per-cycle throughput and lookup tables for efficient leak calculations. Experiments on the Spiking Heidelberg Digits dataset show near-linear scaling of inference time with sparsity levels. Resource counts on the AMD Kria KR260 platform are 740 LUTs, 918 registers, 7 BRAMs, and 24 URAMs, supporting up to 2^17 synapses and 2^10 neurons. The framework is released open-source with NIR integration.

Significance. If the pipeline sustains the claimed throughput, YANA supplies a low-resource, measured FPGA platform that lowers the barrier to neuromorphic hardware experimentation and co-design. The open-source release and direct KR260 measurements (rather than simulation-only results) are concrete strengths that could accelerate algorithmic and hardware development in the field.

major comments (1)

Experiments section: near-linear scaling of inference time is demonstrated solely on the Spiking Heidelberg Digits dataset. The architectural claim that the five-stage pipeline plus preprocessing supports arbitrary SNN topologies with sustained one-event-per-cycle throughput and no buffer overflow therefore rests on an untested extrapolation; additional benchmarks with varied neuron counts, densities, and input rates are required to substantiate generality.

minor comments (2)

Abstract: the phrase 'near-linear scaling' should be accompanied by a quantitative qualifier (e.g., observed slope or R^2) and the exact sparsity ranges tested.
Resource table or text: confirm the exact BRAM/URAM counts and whether they include overhead for the full pipeline.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the single major comment below.

read point-by-point responses

Referee: Experiments section: near-linear scaling of inference time is demonstrated solely on the Spiking Heidelberg Digits dataset. The architectural claim that the five-stage pipeline plus preprocessing supports arbitrary SNN topologies with sustained one-event-per-cycle throughput and no buffer overflow therefore rests on an untested extrapolation; additional benchmarks with varied neuron counts, densities, and input rates are required to substantiate generality.

Authors: We agree that the current experiments are limited to the Spiking Heidelberg Digits dataset and that broader validation would strengthen the generality claims. SHD was selected as it provides a standard, temporally sparse neuromorphic benchmark that directly exercises the sparsity-exploitation features of the pipeline. The architecture itself is designed to be topology-independent: point-to-point connections support arbitrary connectivity graphs, and the five-stage event-driven pipeline with input preprocessing schedules events to sustain one-event-per-cycle throughput without buffer overflow for any topology whose size remains within the hardware bounds (2^10 neurons, 2^17 synapses). In the revised manuscript we will add experiments on additional configurations, including networks with varying neuron counts, connection densities, and input event rates (e.g., synthetic graphs and the NMNIST dataset) to empirically confirm sustained throughput across a wider range of conditions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical hardware measurements on SHD dataset

full rationale

The paper presents a hardware architecture (five-stage event-driven pipeline with input preprocessing) and reports direct measured results on inference-time scaling with sparsity levels using the Spiking Heidelberg Digits dataset. No equations, predictions, or derivations are shown that reduce by construction to fitted parameters, self-definitions, or self-citation chains. The scaling observation is empirical, the resource utilization figures are measured on the AMD Kria platform, and the arbitrary-topology support is stated as a design property without any self-referential proof that collapses to the inputs. This is a standard self-contained hardware-implementation paper with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The design rests on standard digital FPGA principles and existing SNN event-driven models; no free parameters, axioms beyond ordinary hardware assumptions, or new invented entities are introduced.

pith-pipeline@v0.9.0 · 5625 in / 1007 out tokens · 27489 ms · 2026-05-13T18:08:55.685899+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

YANA implements a five-stage, event-driven processing pipeline that fully exploits temporal and spatial sparsity... near-linear scaling of inference time with both spatial and temporal sparsity levels.
IndisputableMonolith/Foundation/Atomicity.lean atomic_tick unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The core operates in discrete timesteps... lookup table (LUT) with precalculated solutions to this term for n_max entries

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

[1]

Advanced Micro Devices, Inc.: PYNQ - Python Productivity for AMD Adaptive Computing Platforms (Feb 2024), https://www.pynq.io/

work page 2024
[2]

et al.: Provable Advantages for Graph Algorithms in Spiking Neural Networks

Aimone, J.B. et al.: Provable Advantages for Graph Algorithms in Spiking Neural Networks. In: Proceedings of the 33rd ACM Symposium on Parallelism in Algo- rithms and Architectures. pp. 35–47. ACM, Virtual Event USA (Jul 2021)

work page 2021
[3]

et al.: Sub-mW Neuromorphic SNN audio processing applications with Rockpool and Xylo (Sep 2022), arXiv:2208.12991 [cs]

Bos, H. et al.: Sub-mW Neuromorphic SNN audio processing applications with Rockpool and Xylo (Sep 2022), arXiv:2208.12991 [cs]

work page arXiv 2022
[4]

et al.: Spiker+: a framework for the generation of efficient Spiking Neural Networks FPGA accelerators for inference at the edge

Carpegna, A. et al.: Spiker+: a framework for the generation of efficient Spiking Neural Networks FPGA accelerators for inference at the edge. IEEE Transactions on Emerging Topics in Computing pp. 1–15 (2024)

work page 2024
[5]

et al.: The Heidelberg Spiking Data Sets for the Systematic Evaluation of Spiking Neural Networks

Cramer, B. et al.: The Heidelberg Spiking Data Sets for the Systematic Evaluation of Spiking Neural Networks. IEEE Transactions on Neural Networks and Learning Systems33(7), 2744–2757 (Jul 2022)

work page 2022
[6]

Davies, M.: Taking Neuromorphic Computing to the Next Level with Loihi 2 (Sep 2021)

work page 2021
[7]

et al.: Comparing Neuromorphic Solutions in Action: Implementing a Bio-Inspired Solution to a Benchmark Classification Task on Three Parallel- Computing Platforms

Diamond, A. et al.: Comparing Neuromorphic Solutions in Action: Implementing a Bio-Inspired Solution to a Benchmark Classification Task on Three Parallel- Computing Platforms. Frontiers in Neuroscience9(Jan 2016)

work page 2016
[8]

et al.: PyTorch lightning (Mar 2019), https://github.com/Lightning- AI/lightning

Falcon, W. et al.: PyTorch lightning (Mar 2019), https://github.com/Lightning- AI/lightning

work page 2019
[9]

et al.: Bottom-Up and Top-Down Approaches for the Design of Neuro- morphic Processing Systems: Tradeoffs and Synergies Between Natural and Arti- ficial Intelligence

Frenkel, C. et al.: Bottom-Up and Top-Down Approaches for the Design of Neuro- morphic Processing Systems: Tradeoffs and Synergies Between Natural and Arti- ficial Intelligence. Proceedings of the IEEE111(6), 623–652 (Jun 2023)

work page 2023
[10]

et al.: The SpiNNaker 2 Processing Element Architecture for Hybrid Digital Neuromorphic Computing (Aug 2022), arXiv:2103.08392 [cs]

Höppner, S. et al.: The SpiNNaker 2 Processing Element Architecture for Hybrid Digital Neuromorphic Computing (Aug 2022), arXiv:2103.08392 [cs]

work page arXiv 2022
[11]

et al.: Neuromorphic computing at scale

Kudithipudi, D. et al.: Neuromorphic computing at scale. Nature637(8047), 801– 812 (Jan 2025)

work page 2025
[12]

et al.: Tonic: event-based datasets and transformations

Lenz, G. et al.: Tonic: event-based datasets and transformations. (Jul 2021), https://doi.org/10.5281/zenodo.5079802

work page doi:10.5281/zenodo.5079802 2021
[13]

et al.: SYNtzulu: A Tiny RISC-V-Controlled SNN Processor for Real- Time Sensor Data Analysis on Low-Power FPGAs

Leone, G. et al.: SYNtzulu: A Tiny RISC-V-Controlled SNN Processor for Real- Time Sensor Data Analysis on Low-Power FPGAs. IEEE Transactions on Circuits and Systems I: Regular Papers72(2), 790–801 (Feb 2025)

work page 2025
[14]

et al.: The road to commercial success for neuromorphic technologies

Muir, D.R. et al.: The road to commercial success for neuromorphic technologies. Nature Communications16(1), 3586 (Apr 2025)

work page 2025
[15]

Pedersen, J.E. et al.: Neuromorphic intermediate representation: A unified instruc- tionsetforinteroperablebrain-inspiredcomputing.NatureCommunications15(1), 8122 (Sep 2024), publisher: Nature Publishing Group

work page 2024
[16]

et al.: Norse - A deep learning library for spiking neural networks (Jan 2021), https://doi.org/10.5281/zenodo.4422025

Pehle, C. et al.: Norse - A deep learning library for spiking neural networks (Jan 2021), https://doi.org/10.5281/zenodo.4422025

work page doi:10.5281/zenodo.4422025 2021
[17]

et al.: DYNAP-SE2: a scalable multi-core dynamic neuromorphic asyn- chronous spiking neural network processor

Richter, O. et al.: DYNAP-SE2: a scalable multi-core dynamic neuromorphic asyn- chronous spiking neural network processor. Neuromorphic Computing and Engi- neering4(1), 014003 (Mar 2024)

work page 2024
[18]

et al.: ModNEF : An Open Source Modular Neuromorphic Emulator for FPGA for Low-Power In-Edge Artificial Intelligence

Saulquin, A. et al.: ModNEF : An Open Source Modular Neuromorphic Emulator for FPGA for Low-Power In-Edge Artificial Intelligence. ACM Transactions on Architecture and Code Optimization p. 3730581 (Apr 2025) YANA: Bridging the Neuromorphic Simulation-to-Hardware Gap 13

work page 2025
[19]

Shrestha,S.B.etal.:EfficientVideoandAudioProcessingwithLoihi2.In:ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). pp. 13481–13485. IEEE, Seoul, Korea, Republic of (Apr 2024)

work page 2024
[20]

et al.: The neurobench framework for benchmarking neuromorphic com- puting algorithms and systems

Yik, J. et al.: The neurobench framework for benchmarking neuromorphic com- puting algorithms and systems. Nature Communications16(1), 1545 (Feb 2025)

work page 2025