Pith. sign in

REVIEW 4 major objections 6 minor 20 references

HiAER-Spike Software-Hardware Reconfigurable Platform for Event-Driven Neuromorphic Computing at Scale

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A reconfigurable FPGA platform for spiking neural networks claims a path to running 160 million neurons and 40 billion synapses at faster than real time, with single-core benchmarks supporting the design.

desk verdict Solid single-core FPGA neuromorphic demo with an honest body, but the abstract sells an unmeasured 160M-neuron projection as fact. read the letter →

arxiv 2602.18072 v1 pith:2HUI7LR2 submitted 2026-02-20 cs.AR cs.AI

Gwenevere Frank , Gopabandhu Hota , Keli Wang , Christopher Deng , Krish Arora , Diana Vins , Abhinav Uppal , Omowuyi Olajide
show 7 more authors
Kenneth Yoshimoto Qingbo Wang Mari Yamaoka Johannes Leugering Stephen Deiss Leif Gibb Gert Cauwenberghs
This is my paper · ORCID
classification cs.ARcs.AI
keywords spikingneuralnetworksneuromorphiccomputingFPGAacceleratoraddress-eventroutingevent-drivenvisionhierarchicalmulticastlarge-scaleSNNinferencereconfigurablehardware
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

HiAER-Spike is a modular, reconfigurable FPGA-based platform for running large spiking neural networks in an event-driven way, meaning neurons compute and communicate only when they spike. The paper's central ambition is scale: the full 40-FPGA system is projected to support 160 million neurons and 40 billion synapses, roughly twice a mouse brain, at faster-than-real-time inference. What is actually demonstrated on a single core is a network of about 110,000 neurons, with software-hardware accuracy agreement within roughly two percentage points on event-based vision benchmarks and per-inference latencies in the microseconds-to-milliseconds range. The authors argue that single-core energy and latency scale linearly with neuron count, and they extrapolate from that trend to the full system. A sympathetic reader would care because the platform is designed as a publicly accessible community resource, pairing a hardware-agnostic Python API with reconfigurable hardware that can be updated as SNN research evolves.

What carries the argument

The carrying mechanism is the hierarchical multicast address-event routing (HiAER) scheme: spike events are encoded as addresses and broadcast through a tree of links—on-chip network, high-speed optical inter-board links, and Ethernet between servers—so traffic is proportional to spiking activity, not network size. Around this, the architecture pairs two memory layers: high-bandwidth memory (HBM) stores the synaptic connectivity table as adjacency lists with base-pointer-plus-row-count entries, while on-chip SRAM (URAM/BRAM) holds membrane potentials and spike registers that are touched every timestep. A two-phase pipeline first enqueues pointers for all fired neurons and axons, then fetches

What would settle it

Run a single network partitioned across two FPGA boards (or two servers) with inter-board spike traffic, and measure per-inference latency and HBM energy at several sizes; if the per-neuron slope increases with the number of routing hops, or if latency grows faster than linearly in neuron count, the 160M-neuron faster-than-real-time projection is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that hierarchical address-event routing makes it possible to run sparse spiking networks efficiently on reconfigurable FPGA hardware and that this efficiency extends to brain scale. Spikes are routed as small address packets through a hierarchy of on-chip, inter-board, and inter-server links, so only the neurons that actually fire consume bandwidth; synaptic connectivity is stored as compact adjacency lists in high-bandwidth memory, with frequently accessed neuron and axon state kept in on-chip SRAM. The demonstrated version of the claim is a single FPGA core executing a 109,615-neuron spiking CNN for DVS gesture recognition, occupying about 86 percent of one cor

Load-bearing premise

The headline capacity rests on the assumption that per-neuron energy and latency measured on one FPGA core (five data points) extend linearly to 1280 cores across 40 FPGAs, with only modest extra delay from inter-core, inter-FPGA, and inter-server routing.

Editorial extensions

If this is right

  • If the linear scaling regressions hold across cores, FPGAs, and servers, the deployed 40-FPGA system reaches the claimed 160M-neuron / 40B-synapse capacity at faster-than-real-time inference.
  • The close software-hardware accuracy match on MNIST, DVS Gesture, CIFAR-10, and Pong validates the conversion pipeline, so researchers can train in a familiar deep-learning setting and deploy to the FPGA without hand-tuning.
  • Because the hardware is reconfigurable, new neuron models and learning rules can be added as new bitstreams, making the platform a community testbed rather than a fixed chip.
  • The event-driven, sparse-activity design means energy and latency per inference are dominated by HBM accesses and scale with active neurons, giving an efficiency profile that is hard to match with dense von Neumann accelerators.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The extrapolation from one core to 1280 cores is the point most worth testing: if inter-FPGA or inter-server routing adds a per-hop penalty that grows with network size, per-neuron cost will rise super-linearly and the headline throughput will not materialize.
  • Because the reported scaling is linear in neuron count rather than in synapse count or spike rate, the trend is most credible for sparse networks with sparse activity; dense connectivity or high firing rates could break the fit.
  • The accuracy gap on DVS Gesture relative to other platforms likely reflects the paper's choice of binarized, downsampled 10-frame inputs rather than a hardware limit; given the same time-step budget and input preprocessing, the gap may shrink.
  • The platform's support for 1 ms spike-timing tracking and online synapse writes suggests a natural next step—testing online learning rules like STDP across the full system—which would be a direct evaluation of the routing hierarchy under sustained traffic.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents HiAER-Spike, an FPGA-based, event-driven neuromorphic platform assembled at the San Diego Supercomputer Center, with hierarchical address-event routing, HBM-backed synaptic storage, a Python API, and access through the Neuroscience Gateway. The demonstrated results are from a single core on a single FPGA: binarized MNIST MLP/LeNet variants, DVS Gesture spiking CNNs, a CIFAR-10 spiking CNN, and a Pong DQN. Software–hardware accuracies match exactly on four MNIST configurations and within roughly two percentage points on the larger DVS and CIFAR-10 models. The paper also presents linear scaling regressions for HBM-energy and latency versus neuron count, and uses these to project a full-system capacity of 160 million neurons and 40 billion synapses at faster-than-real-time speed.

Significance. If the full-scale system were realized and measured, it would be a unique publicly accessible FPGA-based neuromorphic resource: 40 FPGAs, reconfigurable logic, and a community-facing Python interface would distinguish it from fixed ASIC efforts such as Loihi and TrueNorth. The single-core evidence is credible and carefully executed: the authors validated conversions against SpikingJelly, deliberately matched neuron update order and threshold semantics, and report exact software–hardware agreement on several MNIST models. The open availability of the hs_api code and the explicit n=5 caveat on the scaling regressions are also strengths. However, the central scale claim—160M neurons, 40B synapses, faster-than-real-time—is not measured; it is an extrapolation from a single-core linear fit, and the reported energy is an HBM-access proxy rather than a measured system energy. The significance of the paper therefore depends on whether the projected scaling is validated or clearly reclassified as a design goal.

major comments (4)
  1. [Abstract, §3, §6] The headline claim of 160 million neurons, 40 billion synapses, and faster-than-real-time execution is not supported by the presented measurements. The largest deployed network has 109,615 neurons on one core (§6), and the scaling regressions (Energy = 0.0294x − 30.293, Latency = 0.0658x − 53.031, R² = 0.994/0.995, n = 5) are fit over single-core neuron counts. The full system would use 1280 cores; the relevant unknowns are inter-core, inter-FPGA, and inter-server routing latency, partitioning behavior, and spike loss. The paper's own §6 statement that the authors 'anticipate only modest deviations from these linear trends' is an assumption, not a result. I recommend either removing or explicitly labeling the headline as projected capacity, or providing even a small multi-core/multi-FPGA scaling measurement.
  2. [§6, Tables 2–4] The energy metric is an HBM-access proxy, not a measured system energy. The text states: 'energy consumption was approximated by the product of the energy cost of a single HBM access and the number of HBM accesses performed during an inference.' Yet Tables 3 and 4 compare this 'HBM Energy' directly with Loihi, SpiNNaker, and TrueNorth energies and the paper claims 'superior energy and latency metrics.' Without stating the per-access energy constant or reporting wall-power measurements, these cross-platform comparisons are not apples-to-apples. Report HBM-access counts as a separate metric, or measure total energy.
  3. [§6, Tables 3–4] The cross-platform benchmark comparisons are uncontrolled. HiAER-Spike MNIST results use ANN (binary) neurons rather than spiking neurons; DVS Gesture inputs are binarized and downsampled to (63,63) or (90,90) rather than the full (128,128); and timestep counts differ from the compared systems. The authors acknowledge these differences, but the tables still present the numbers as comparative performance. Since the paper's goal is platform demonstration rather than accuracy leadership, I suggest recasting these tables as reference points with explicit caveats in the table captions, rather than as competitive benchmarks.
  4. [§3, A.3] The network partitioning and resource allocation algorithm, cited as reference [10], is load-bearing for the full-system scale claim, yet it is neither described nor experimentally validated in this manuscript. Section A.3 says partitioning 'follows the scheme described in [10],' but no partitioning experiment is reported. The authors should provide at least a high-level description of the algorithm and a small multi-core partition test (e.g., splitting a single network across 2, 4, or 8 cores and measuring latency/energy), or state explicitly that multi-core partitioning is future work and remove the implication that it is operational.
minor comments (6)
  1. [Abstract, §3] The abstract says the system is 'assembled at the UC San Diego Supercomputer Center,' while §3 states 'This paper presents initial benchmarks using only a single operational core on one FPGA.' Please harmonize these statements so the reader immediately understands that the full 40-FPGA system is not yet benchmarked.
  2. [Abstract, §6] 'Faster than real time' is used without a definition. Specify the real-time reference (e.g., wall-clock time of the input stream, biological real time, or a standard SNN simulation time budget).
  3. [Table 2, §6] The row for the largest DVS Gesture model reports 109,615 neurons and states it occupies ~86% of a single-core system, but the core's neuron and synapse capacity limits are not given. State the single-core limits explicitly.
  4. [Fig. 8] In the simulator code, the line 'perturbation (perturbation | 1)' appears to be missing an assignment; it should likely be 'perturbation = perturbation | 1' or 'perturbation |= 1'.
  5. [References] Reference [10] is incomplete ('15(222)'); it needs full venue, year, and page/article information.
  6. [§6] The text mentions a resized MNIST variant with a 0.01% accuracy discrepancy that is 'not included in Table 2.' Either include the configuration details or omit the sentence, since the reader cannot verify the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmarks are externally validated and the scale-up claim is a hedged extrapolation, not a fitted input presented as an independent prediction.

full rationale

The paper makes no derived claim that is equivalent to an input by construction. Accuracy results are checked against external PyTorch/SpikingJelly software models with deliberate matching of neuron update order, so the hardware/software agreement is not forced by fitting. The 160M-neuron / 40B-synapse figure is stated as a design aim ('We aim for each FPGA to support up to four million neurons and one billion synapses, yielding a total capacity of up to 160 million neurons and 40 billion synapses'), not as a quantity inferred from the single-core data. The §6 scaling regressions (Energy = 0.0294x−30.293, Latency = 0.0658x−53.031) are fits to the paper's own single-core measurements, but they are explicitly labeled as limited observations ('these observations are limited to a small sample size (n=5)'), and the multi-core statement is a hedged anticipation ('we anticipate only modest deviations from these linear trends'), not a prediction disguised as independent evidence. The self-citations [7], [8], and [10] describe the HiAER routing and partitioning architecture as design background; no result here reduces to an unverified self-citation chain or imported uniqueness theorem. No ansatz is smuggled in through citation, and no known result is merely renamed. The scale-up extrapolation is a correctness/validation risk rather than a circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central projection rests on three domain assumptions (routing scales, simulator fidelity, HBM packing) and two fitted/assumed constants (HBM energy per access, linear scaling slopes). No new physical entities are introduced. The single-core benchmark results themselves do not depend on these assumptions.

free parameters (2)
  • Energy per HBM access constant
    HBM energy per inference is computed as the number of HBM accesses reported by the FPGA multiplied by an assumed per-access energy cost; the constant is not given, so the energy numbers are not independently checkable (Table 2 notes, §6).
  • Energy/latency linear scaling slope and intercept (DVS Gesture) = Energy = 0.0294x − 30.293 µJ; Latency = 0.0658x − 53.031 µs (x = neuron count)
    Fitted to n=5 single-core data points (R²=0.994 and 0.995) and then used to argue that full-system scaling will follow the same linear trend with 'modest deviations' (§6).
assumptions (3)
  • domain assumption The HiAER multicast routing and partitioning schemes of refs [7], [8], [10] scale to 1280 cores across 40 FPGAs with only modest added latency.
    The paper's capacity and throughput projection depends on this, but no multi-FPGA measurements are presented (§3, §6).
  • domain assumption The Python simulator faithfully emulates FPGA hardware for arbitrary networks.
    The simulator is validated only on the small benchmark set; it is used as a stand-in for hardware in the software path (§5.3).
  • domain assumption 8 GB of HBM per FPGA can store ~1 billion synapses under the slot-aligned adjacency-list format.
    The 1B-synapse capacity target requires the 16-slot HBM alignment rule to achieve sufficient packing density; the packing efficiency is not quantified (§4, §3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of HiAER-Spike Software-Hardware Reconfigurable Platform for Event-Driven Neuromorphic Computing at Scale." pith.science (2026). https://pith.science/paper/2HUI7LR2

@misc{pith2026260218072,
  author       = {Pith},
  title        = {Pith review of: HiAER-Spike Software-Hardware Reconfigurable Platform for Event-Driven Neuromorphic Computing at Scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2HUI7LR2}},
  note         = {Machine review of arXiv:2602.18072}
}
read the original abstract

In this work, we present HiAER-Spike, a modular, reconfigurable, event-driven neuromorphic computing platform designed to execute large spiking neural networks with up to 160 million neurons and 40 billion synapses - roughly twice the neurons of a mouse brain at faster than real time. This system, assembled at the UC San Diego Supercomputer Center, comprises a co-designed hard- and software stack that is optimized for run-time massively parallel processing and hierarchical address-event routing (HiAER) of spikes while promoting memory-efficient network storage and execution. The architecture efficiently handles both sparse connectivity and sparse activity for robust and low-latency event-driven inference for both edge and cloud computing. A Python programming interface to HiAER-Spike, agnostic to hardware-level detail, shields the user from complexity in the configuration and execution of general spiking neural networks with minimal constraints in topology. The system is made easily available over a web portal for use by the wider community. In the following, we provide an overview of the hard- and software stack, explain the underlying design principles, demonstrate some of the system's capabilities and solicit feedback from the broader neuromorphic community. Examples are shown demonstrating HiAER-Spike's capabilities for event-driven vision on benchmark CIFAR-10, DVS event-based gesture, MNIST, and Pong tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 1 linked inside Pith

  1. [10]

    Accessed 2023-04-03 18

    Mysore, N., Hota, G., Deiss, S.R., Pedroni, B.U., Cauwenberghs, G.: Hierarchical network connectivity and partitioning for reconfigurable large-scale neuromorphic systems15(222). Accessed 2023-04-03 18

    2023

  2. [1]

    Neural networks 121, 294–307 (2020)

    Deng, L., Wu, Y., Hu, X., Liang, L., Ding, Y., Li, G., Zhao, G., Li, P., Xie, Y.: Rethinking the performance comparison between snns and anns. Neural networks 121, 294–307 (2020)

    2020

  3. [2]

    In: Practice and Experience in Advanced Research Computing

    Sivagnanam, S., Yoshimoto, K., Carnevale, T., Nadeau, D., Kandes, M., Petersen, T., Truong, D., Martinez, R., Delorme, A., Makeig, S., Majumdar, A.: Neuro- science gateway enabling large scale modeling and data processing in neuroscience research. In: Practice and Experience in Advanced Research Computing. PEARC ’20, pp. 510–513. Association for Computing...

    arXiv 2023

  4. [3]

    In: Proceedings of the IEEE 2012 Custom Integrated Circuits Conference, pp

    Painkras, E., Plana, L.A., Garside, J., Temple, S., Davidson, S., Pepper, J., Clark, D., Patterson, C., Furber, S.: Spinnaker: A multi-core system-on-chip for massively-parallel neural net simulation. In: Proceedings of the IEEE 2012 Custom Integrated Circuits Conference, pp. 1–4 (2012). IEEE

    2012

  5. [4]

    Intel Labs’ Loihi2, 1–7 (2021)

    Davies, M.,et al.: Taking neuromorphic computing to the next level with loihi2. Intel Labs’ Loihi2, 1–7 (2021)

    2021

  6. [5]

    Frontiers in Neuroscience16, 795876 (2022)

    Pehle, C., Billaudelle, S., Cramer, B., Kaiser, J., Schreiber, K., Stradmann, Y., Weis, J., Leibfried, A., M¨ uller, E., Schemmel, J.: The brainscales-2 accelerated neuromorphic system with hybrid plasticity. Frontiers in Neuroscience16, 795876 (2022)

    2022

  7. [6]

    Science382(6668), 329–335 (2023)

    Modha, D.S., Akopyan, F., Andreopoulos, A., Appuswamy, R., Arthur, J.V., Cassidy, A.S., Datta, P., DeBole, M.V., Esser, S.K., Otero, C.O.,et al.: Neural inference at the frontier of energy, space, and time. Science382(6668), 329–335 (2023)

    2023

  8. [7]

    IEEE trans- actions on neural networks and learning systems28(10), 2408–2422 (2016)

    Park, J., Yu, T., Joshi, S., Maier, C., Cauwenberghs, G.: Hierarchical address event routing for reconfigurable large-scale neuromorphic systems. IEEE trans- actions on neural networks and learning systems28(10), 2408–2422 (2016)

    2016

Show all 20 references
  1. [8]

    In: 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pp

    Hota, G., Mysore, N., Deiss, S., Pedroni, B., Cauwenberghs, G.: Hierarchical multicast network-on-chip for scalable reconfigurable neuromorphic systems. In: 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 481– 485 (2022). https://doi.org/10.1109/ISCAS487...

    2022

  2. [9]

    Proceedings of the National Academy of Sciences103(32), 12138–12143 (2006)

    Herculano-Houzel, S., Mota, B., Lent, R.: Cellular scaling rules for rodent brains. Proceedings of the National Academy of Sciences103(32), 12138–12143 (2006)

    2006

  3. [11]

    In: 2019 26th IEEE International Conference on Electronics, Circuits and Systems (ICECS), pp

    Kulkarni, S.R., Kadetotad, D.V., Yin, S., Seo, J.-S., Rajendran, B.: Neuromor- phic hardware accelerator for snn inference based on stt-ram crossbar arrays. In: 2019 26th IEEE International Conference on Electronics, Circuits and Systems (ICECS), pp. 438–441 (2019). IEEE

    2019

  4. [12]

    IEEE Journal on Exploratory Solid- State Computational Devices and Circuits6(1), 36–44 (2020)

    Correll, J.M., Bothra, V., Cai, F., Lim, Y., Lee, S.H., Lee, S., Lu, W.D., Zhang, Z., Flynn, M.P.: A fully integrated reprogrammable cmos-rram compute-in-memory coprocessor for neuromorphic applications. IEEE Journal on Exploratory Solid- State Computational Devices and Circui...

    2020

  5. [13]

    Frontiers in neuroscience13, 357 (2019)

    Pedroni, B.U., Joshi, S., Deiss, S.R., Sheik, S., Detorakis, G., Paul, S., Augustine, C., Neftci, E.O., Cauwenberghs, G.: Memory-efficient synaptic connectivity for spike-timing-dependent plasticity. Frontiers in neuroscience13, 357 (2019)

    2019

  6. [14]

    arXiv preprint arXiv:2309.16795 (2023)

    Lenz, G., Orchard, G., Sheik, S.: Ultra-low-power image classification on neuro- morphic hardware. arXiv preprint arXiv:2309.16795 (2023)

    2023 arXiv

  7. [15]

    In: 2015 International Joint Conference on Neural Networks (IJCNN), pp

    Stromatias, E., Neil, D., Galluppi, F., Pfeiffer, M., Liu, S.-C., Furber, S.: Scalable energy-efficient, low-latency implementations of trained spiking deep belief net- works on spinnaker. In: 2015 International Joint Conference on Neural Networks (IJCNN), pp. 1–8 (2015). IEEE

    2015

  8. [16]

    Advances in neural information processing systems28(2015)

    Esser, S.K., Appuswamy, R., Merolla, P., Arthur, J.V., Modha, D.S.: Back- propagation for energy-efficient neuromorphic computing. Advances in neural information processing systems28(2015)

    2015

  9. [17]

    In: 2020 International Joint Conference on Neural Networks (IJCNN), pp

    Massa, R., Marchisio, A., Martina, M., Shafique, M.: An efficient spiking neural network for recognizing gestures with a dvs camera on the loihi neuromorphic processor. In: 2020 International Joint Conference on Neural Networks (IJCNN), pp. 1–9 (2020). IEEE

    2020

  10. [18]

    In: 2025 Neuro Inspired Computational Elements (NICE), pp

    Arfa, S., Vogginger, B., Liu, C., Partzsch, J., Sch¨ one, M., Mayr, C.: Efficient deployment of spiking neural networks on spinnaker2 for dvs gesture recogni- tion using neuromorphic intermediate representation. In: 2025 Neuro Inspired Computational Elements (NICE), pp. 1–8 (2...

    2025

  11. [19]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Amir, A., Taba, B., Berg, D., Melano, T., McKinstry, J., Di Nolfo, C., Nayak, T., Andreopoulos, A., Garreau, G., Mendoza, M.,et al.: A low power, fully event- based gesture recognition system. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p...

    2017

  12. [20]

    alpha" axon with synapses to the

    Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G.,et al.: Human-level control through deep reinforcement learning. nature518(7540), 529–533 (2015) 19 A Supplementary Information A.1 API De...

    2015

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.