Pith. sign in

REVIEW 3 major objections 9 minor 89 references

The SpiNNaker2 chip: a many-core platform for flexible and scalable brain-inspired computing

T0 review · 3 major / 9 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read SpiNNaker2 is a many-core chip that runs both spiking networks and deep networks on the same flexible substrate, with low idle power so sparse and event-based work stays efficient.

desk verdict Solid measured silicon paper on a real 152-PE hybrid chip; the headline TOPS numbers are SRAM peaks and the body already admits DRAM kills sustained DNN throughput. read the letter →

arxiv 2607.24396 v1 pith:DEARRAMT submitted 2026-07-27 cs.ET cs.AIcs.ARcs.DC

Stefan Scholze , Johannes Partzsch , Sebastian Höppner , Florian Kelber , Andreas Dixius , Marco Stolba , Sirine Arfa , Marc Berthel
show 22 more authors
Georg Ellguth Jim Garside Hector A. Gonzalez Stephan Hartmann Thomas Kiel-Hocker Dongwei Hu Matthias Jobst Khaleelulla Khan Nazeer Tim Langer Chen Liu Gengting Liu Matthias Lohrmann Mantas Mikaitis Felix Neumärker Amirhossein Rostami Stefan Schiefer Tilo Schubert Delong Shang Bernhard Vogginger Yexin Yan Steve Furber Christian Mayr
This is my paper
classification cs.ETcs.AIcs.ARcs.DC
keywords SpiNNaker2neuromorphiccomputingmany-corespikingneuralnetworksdeepevent-basedcommunicationDVFSINT8inference
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

The paper presents SpiNNaker2, a manufactured many-core chip meant to close the gap between neuromorphic hardware and mainstream deep learning. It puts 152 small programmable cores on one die, each with an ARM processor plus accelerators for matrix math, convolutions, exponentials and rounding, and ties them together with an extended multicast event router and chip-to-chip links. Measured silicon reaches up to 4.5 TOPS in high-performance INT8 mode and up to 2.7 TOPS/W in high-efficiency mode, can host more than 150,000 spiking neurons and more than 1.8 billion synaptic events per second at a 1 ms step, and keeps baseline power under 250 mW. That low floor, plus per-core voltage and frequency scaling, is meant to turn algorithmic sparsity into real energy savings. The authors argue this makes the chip a universal platform for brain-inspired computing and for hybrids that mix spikes, events and conventional deep layers, from single-chip robots up to multi-million-core machines.

What carries the argument

The processing element: an ARM M4F plus local SRAM and dedicated accelerators (matrix/convolution MLA, exp/log, RNG, rounding), grouped in quads with fine-grained DVFS, and linked by an extended SpiNNaker multicast event router plus a dual NoC. This combination carries the claim that software flexibility and event communication can still hit competitive efficiency.

What would settle it

Run the same multi-layer DNN and large SNN end-to-end on the chip with all tensors forced through the LPDDR4 interfaces (and, for scale, multi-chip event links), measure sustained TOPS/W and synaptic-event rate, and compare those numbers to the on-SRAM tile results and to a dense GPU baseline under identical batch and real-time constraints.

Watch

Extended reading notes

Core claim

On measured silicon, SpiNNaker2 shows that one programmable many-core design can serve neuromorphic, deep-network and novel event-based workloads at once: INT8 deep layers reach up to 4.5 TOPS and 2.7 TOPS/W, spiking nets support more than 150,000 neurons and more than 1.8 billion synaptic events per second at 1 ms, and baseline power stays under 250 mW so varying and sparse loads remain efficient.

Load-bearing premise

The efficiency story rests on workloads whose weights and state largely fit in on-chip SRAM; once large multi-layer nets must stream through DRAM and multi-chip links, that advantage may shrink.

Editorial extensions

If this is right

  • Sparse and event-based deep models can turn reduced activity directly into lower energy via DVFS and sub-250 mW idle power.
  • Hybrid spike–ANN and flexible-payload learning rules (for example event-based backprop) become practical on the same hardware as ordinary INT8 inference.
  • Single chips can act as standalone sensor–actuator controllers while the same architecture scales toward multi-million-core systems.
  • On-chip scheduling of tiled DNN layers can keep worker utilization high when data stays local, guiding when to pipeline or depth-first map larger models across chips.

Reading between the lines

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

  • If DRAM bandwidth remains the limiter for dense layers, the platform’s comparative edge will concentrate on irregular, sparse, or streaming workloads rather than bulk dense training or inference.
  • Per-core power modes and multicast events suggest natural fits for mixture-of-experts and state-space models that activate only subsets of compute over time.
  • A fair multi-generation SNN capacity comparison will need a shared software stack; until then, claimed gains over the predecessor stay partly software-dependent.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 9 minor

Summary. The manuscript presents the SpiNNaker2 chip, a 152-processing-element many-core in GlobalFoundries 22FDX with per-PE ARM M4F cores, a machine-learning accelerator (16x4 MAC array), numerical/rounding/RNG accelerators, fine-grained DVFS with adaptive body biasing, an extended SpiNNaker multicast event router, short- and long-range chip-to-chip links, two LPDDR4 interfaces, and Gbit Ethernet. The paper documents the architecture and physical implementation and reports measured results: baseline power in several sleep/active configurations (Tab. 2), numerical accelerator speedups (Fig. 7), MLA vs. CMSIS throughput and efficiency (Fig. 8, Tab. 3), per-core SNN capacity (Fig. 9), a DVS-Gesture SNN with automatic DVFS giving 28% energy savings (Tab. 4), full-layer DNN execution with DRAM streaming under OctopuScheduler (Tab. 5, Fig. 11), and brief reports of event-based algorithms (E-prop, EventProp, EGRU) and a robotics demonstrator. The central claim is that SpiNNaker2 is a flexible, scalable platform bridging neuromorphic and deep-network workloads. The claims are backed by chip measurements, with caveats mostly disclosed in the body.

Significance. If the numbers stand — and they are measurements on fabricated silicon — this is a substantial contribution: a 152-core 22 nm FDSOI neuromorphic many-core with per-PE ML and numerical accelerators, measured per-core DVFS with a demonstrated 28% energy saving on a real SNN workload, and a routing fabric already deployed at 5-million-core scale. The paper is unusually candid about its own limits (Sec. V explicitly concedes DRAM-bound DNN scalability), ships reproducible benchmark code, and documents accelerators, router internals, and link protocols to a level that makes the platform usable and assessable by others. It is the definitive reference for an architecture the community will build on, and it honestly occupies the under-served middle ground between SNN-only neuromorphic chips and DNN-only accelerators.

major comments (3)
  1. [IV-C, Fig. 8, Tab. 3] §IV-C, Fig. 8, and Tab. 3: the headline 4.563 TOPS / 2.77 TOPS/W figures are not full-chip measurements. Fig. 8's caption defines 'full utilization' as 'a scenario if all 152 PEs would be used with the same configuration' — a linear extrapolation of per-tile measurements assuming all weights/inputs reside in the 128 kB PE SRAM and no NoC/DRAM contention under concurrent MLA prefetch. Yet Tab. 3 footnotes these values with '*' denoting 'measured values'. Please relabel them as projected/extrapolated (per-PE measured, full-chip scaled), and add a qualifier to the abstract's 'up to 4.5 TOPS' claim. The paper's own §IV-E/Tab. 5/Fig. 11c bound the realistic regime: MM2 sustains ~0.1 TOPS at the same clock (24.5 ms, 98% of time in memory transfers), a >20x gap. Sec. V discloses this candidly, but the qualifier belongs with the headline numbers, not only in the conclusion.
  2. [IV-D, Fig. 9] Abstract and §IV-D: the claims of '>150000 neurons and >1.8 billion synaptic events/s' appear to be per-core measurements (Fig. 9: ~12M synaptic events/s/core at 1024 neurons) multiplied linearly by 152 cores. The provenance should be stated in the abstract or §IV-D, together with whether the event router and spike FIFO paths were characterized at the implied full-chip aggregate spike rate, or whether this is likewise an extrapolation assuming no router saturation and perfectly distributed traffic. Given the router's 1843.2 Gbit/s peak this is plausible, but the reader should not have to reconstruct the arithmetic.
  3. [IV-E, Tab. 5, supplement VI-F] Tab. 5 vs. supplement §VI-F: the cross-platform DNN comparison uses mismatched batch sizes — the A100/torchao FC-layer numbers are for batch 17 (kernel limitation), while OctopuScheduler runs FC layers at batch 4m and Tab. 9a shows batch-1 FC inputs — yet Tab. 5 reports per-layer time and energy without per-inference normalization. The GPU power figures (107-133 W) also appear to be board-level against SpiNNaker2 chip-level power. Since energy ordering between platforms can flip under normalization, please either normalize per inference in the main text or add an explicit caveat to Tab. 5's caption.
minor comments (9)
  1. [Tab. 3] Tab. 3, GAP9 row: 0.151 TOPS / 0.64 W implies 0.24 TOPS/W, but the table reports 0.34. Either the efficiency figure comes from a different operating point in [8] or there is an arithmetic error; please reconcile.
  2. [IV-E] §IV-E, first paragraph: 'the same 6 DNN layers from Section IV-E' is a self-reference; this should be Section IV-C.
  3. [Abstract, IV-A] Abstract and Tab. 2: the '<250 mW baseline power' claim should name its configuration — Tab. 2 shows 75.2 mW (PEs off) to 564.7 mW (PEs sleep @0.8 V); presumably the 235.4 mW sleep@0.5 V operating point is meant.
  4. [IV-B, Fig. 7] Fig. 7 vs. text: the text quotes 5.3x/4.8x for exp/log and 4.4x-5.7x for the optimized variant, while the figure shows five bars labeled 5.7x/5.1x/5.0x/4.4x/4.4x with only some marked '*'. The mapping between bars and implementation variants is ambiguous; please clarify the legend.
  5. [IV-D, Tab. 4] §IV-D / Tab. 4: the 92.04% chip accuracy is measured on a 1 s (1000-step) window while the 94.0% PyTorch reference uses full ~6 s gestures; these are not directly comparable and the text should say so explicitly next to Tab. 4, not only in the narrative.
  6. [IV-F, supplement VI-G] §IV-F and supplement §VI-G: the '8x less energy' E-prop claim depends on the 'effective power' normalization (Eq. 1-2) that scales baseline power by 7.9% utilization; unnormalized, 0.29 W x 500 h exceeds the V100's energy. Please report the raw measured energy alongside, and note the batch-size asymmetry (1 vs. 100).
  7. [various] Typos: Fig. 4 label 'Perihery' -> 'Periphery'; Tab. 6 '1000 Mhz' -> 'MHz'; §IV-A 'Midst coremark execution' reads awkwardly; supplement §VI-E.2 'the maximum number of timesteps is then interpolated' presumably means synaptic events/s.
  8. [IV-A] §IV-A: the power-measurement methodology would benefit from one or two sentences on instrument, measurement bandwidth/averaging, and which rails are included in 'total' (periphery, IO, PLL are listed but DRAM and board components are not), since several headline efficiency numbers rest on these measurements.
  9. [IV-D, supplement VI-E] §IV-D and supplement VI-E.1: the possibility of PEs drifting out of phase (missed timer interrupts) is noted and logged; for the DVS-Gesture experiment it would strengthen the real-time claim to report how many such incidents occurred during the measured runs.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: load-bearing claims are silicon measurements and software benchmarks, not derivations that re-encode inputs as predictions.

full rationale

SpiNNaker2 is a hardware/systems paper whose central claims (INT8 TOPS/TOPS/W, SNN neuron and synaptic-event capacity, baseline power <250 mW, event-based algorithm energy comparisons) are empirical outputs from a fabricated 22FDX chip and accompanying software stacks. There is no first-principles derivation chain in which a fitted parameter, self-defined quantity, or author-only uniqueness theorem is renamed as a prediction. Self-citations (SpiNNaker routing fabric, DVFS [20], numerical/MLA accelerators, py-spinnaker2, OctopuScheduler, prior EGRU/EventProp/E-prop demos) supply component background and prior micro-architecture; they do not force the measured chip metrics. The Fig. 8 “full utilization” scaling and the abstract’s “up to” peak numbers are conventional peak-vs-sustained reporting (with DRAM-bound sustained results disclosed in §IV-E/Fig. 11 and §V), not circular reduction of fit to forecast. No self-definitional loop, fitted-input-as-prediction, load-bearing uniqueness import, or renaming of a known result is present. Score 0 is the appropriate honest finding.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

As a digital systems/measurement paper, the claim rests on standard VLSI and computer-architecture assumptions plus design choices (voltages, clocks, SRAM-first mapping), not on new physical entities. The ledger records those background assumptions and the few free design/operating parameters that gate the headline efficiency numbers.

free parameters (4)
  • PE voltage/frequency pairs (0.5 V @ 150 MHz vs 0.8 V @ 300 MHz) = 0.5V/150MHz and 0.8V/300MHz defaults
    Headline TOPS and TOPS/W are defined at these two characterized operating points chosen by the design; other PVT corners are not the reported peaks.
  • Auto-PL spike-count threshold for DVFS in DVS-Gesture SNN = user-defined threshold from profiling (value not fixed universally)
    Energy saving in Auto PL depends on a user-defined received-spike threshold fitted from profiling “time done” vs spikes (§IV-D).
  • Neurons-per-core and connectivity assumptions in SNN capacity = up to ~1024 neurons/core in capacity sweep; app uses hand-tuned partition (Tab. 7)
    “>150k neurons” and synaptic-event rates depend on chosen neuron model simplicity, 16-bit synapses in SRAM, 1 ms tick, and mapping density (Fig. 9, supplement).
  • MLA utilization / tiling choices in OctopuScheduler = e.g. 78% MAC util for MM full-chip case
    Reported 4.5 TOPS and layer energies depend on partition shapes that maximize MAC-array use and on keeping tensors local when claiming peak efficiency.
assumptions (5)
  • domain assumption Standard synchronous digital CMOS / FD-SOI timing and power models with adaptive body bias guarantee speed across characterized PVT corners.
    Physical design section relies on ABX-aware implementation and STA per macro (§III-C, III-D).
  • domain assumption Discrete-time Euler software neuron updates with timer-driven PE wakeups adequately represent the SNN workloads claimed.
    SNN results assume 1 ms (or 3 ms) ticks and accept possible phase slip under overload (§IV-D, supplement synchronization).
  • domain assumption INT8 MAC throughput counted as 2 OPs per MAC is the right comparison basis versus other edge accelerators in Tab. 3.
    Standard in the accelerator survey literature the paper cites; still a convention that shapes TOPS/W rankings.
  • domain assumption Hexagonal multicast event fabric plus small payloads is an efficient match for sparse brain-like and event-based traffic at scale.
    Design philosophy inherited from SpiNNaker (§II); underpins scalability claims toward 10M cores.
  • standard math Arithmetic and bus protocols of ARM M4F, AHB/APB, LPDDR4, Ethernet/UDT behave as specified by vendors/standards.
    Background engineering standards; not re-proved in the paper.
invented entities (2)
  • SpiNNaker2 chip (152-PE GALS many-core with MLA, numerical accel, extended event router) independent evidence
    purpose: Concrete hardware substrate realizing the flexible neuromorphic+DNN platform claims
    The primary engineered artifact; not a postulated particle/force but a new physical device whose properties are measured in this work.
  • OctopuScheduler on-chip DNN tiling/scheduling framework independent evidence
    purpose: Map multi-layer DNNs across PEs with DRAM movement accounted for
    Software entity introduced/used to obtain full-layer DNN numbers; details largely in companion citations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The SpiNNaker2 chip: a many-core platform for flexible and scalable brain-inspired computing." pith.science (2026). https://pith.science/paper/DEARRAMT

@misc{pith2026260724396,
  author       = {Pith},
  title        = {Pith review of: The SpiNNaker2 chip: a many-core platform for flexible and scalable brain-inspired computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DEARRAMT}},
  note         = {Machine review of arXiv:2607.24396}
}
read the original abstract

In deep learning, efficiency gets more and more important to compensate for the ongoing growth in model sizes and applications. Neuromorphic hardware has long been advocated as an upcoming alternative to deep networks, taking inspiration from the brain for achieving unprecedented energy efficiency. However, demonstrations of these gains only recently began to grow in complexity and real-world applicability. With SpiNNaker2, we present a chip that bridges the gap between deep networks and neuromorphic computing and allows for flexible exploration of computing approaches that combine both worlds. It features 152 processing elements equipped with an ARM M4F processor and dedicated accelerators, an extended SpiNNaker routing fabric for scalable event-based communication and a range of external interfaces for system integration, including Gbit Ethernet and an LPDDR4 memory interface. We demonstrate performance and efficiency of the SpiNNaker2 chip for neuromorphic and deep network workloads, as well as novel event-based computing approaches. For deep network workloads, the chip achieves up to 4.5 TOPS in high performance mode and up to 2.7 TOPS/W efficiency in high efficiency mode for INT8 workloads. The chip supports spiking neural networks with >150000 neurons and >1.8 billion synaptic events/s when simulated with a 1 ms time step. Its low baseline power of less than 250 mW allows for efficiency even under varying workload conditions, allowing to explore sparse and event-based modes of computation. All this demonstrates the chip's capabilities as a universal hardware platform for scalable brain-inspired computing and its combinations with mainstream deep network approaches.

Figures

Figures reproduced from arXiv: 2607.24396 by the authors.

Figure 1
Figure 1. FIGURE 1: Hexagonal SpiNNaker communication fabric. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIGURE 2: SpiNNaker2 chip layout with components [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIGURE 3: SpiNNaker2 processing element (PE) compo [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 6
Figure 6. Figure 6: FIGURE 6: SpiNNaker2 communication interfaces [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 5
Figure 5. Figure 5: FIGURE 5: SpiNNaker2 event router [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 8
Figure 8. Figure 8: FIGURE 8: Performance and energy efficiency for 6 DNN [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 7
Figure 7. Figure 7: FIGURE 7: Time measurements for the numerical accelera [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: FIGURE 9: Maximum synaptic events per second per core [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: FIGURE 10: SNN for DVS-Gesture recognition using dynamic power management. Left: NIR model of deep SNN and [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: FIGURE 11: Execution times of layers based on (a) 2D con [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 13
Figure 13. Figure 13: FIGURE 13: Internal structure of the SpiNNaker2 event [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

89 extracted references · 7 linked inside Pith

  1. [1]

    Benchmarking neuromorphic hardware and its energy expenditure,

    C. Ostrau, C. Klarhorst, M. Thies, and U. R ¨uckert, “Benchmarking neuromorphic hardware and its energy expenditure,”Frontiers in Neuroscience, Section Neuromorphic Engineering, vol. 16, 2022

    2022

  2. [2]

    Neuromorphic hardware for sustainable ai data centers,

    B. V ogginger et al., “Neuromorphic hardware for sustainable ai data centers,” 2024

    2024

  3. [3]

    Exponentially faster language mod- elling,

    P. Belcak and R. Wattenhofer, “Exponentially faster language mod- elling,” 2023. https://arxiv.org/abs/2311.10770

    arXiv 2023

  4. [4]

    Mixture of a million experts,

    X. O. He, “Mixture of a million experts,” 2024. https://arxiv.org/abs/ 2407.04153

    arXiv 2024

  5. [5]

    Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models,

    M. Sch ¨oneet al., “Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models,” 2024. https://arxiv.org/ abs/2404.18508. VOLUME , 13 :

    arXiv 2024

  6. [6]

    Sambanova SN10 RDU: A 7nm dataflow architecture to accelerate software 2.0,

    R. Prabhakar, S. Jairath, and J. L. Shin, “Sambanova SN10 RDU: A 7nm dataflow architecture to accelerate software 2.0,” inIEEE ISSCC, vol. 65, pp. 350–352, 2022

    2022

  7. [7]

    The wormhole ai training processor,

    D. Ignjatovi ´c, D. W. Bailey, and L. Baji ´c, “The wormhole ai training processor,” in2022 IEEE International Solid-State Circuits Conference (ISSCC), vol. 65, pp. 356–358, 2022

    2022

  8. [8]

    Lincoln AI Computing Survey (LAICS) Update,

    A. Reutheret al., “Lincoln AI Computing Survey (LAICS) Update,” inHPEC 2023, pp. 1–7, 2023. ISSN: 2643-1971

    2023

Show all 89 references
  1. [9]

    Neuromorphic hardware in the loop: Training a deep spiking network on the brainscales wafer-scale system,

    S. Schmitt et al., “Neuromorphic hardware in the loop: Training a deep spiking network on the brainscales wafer-scale system,” in International Joint Conference on Neural Networks (IJCNN), 2017

    2017

  2. [10]

    Reckon: A 28nm sub-mm2 task-agnostic spiking recurrent neural network processor enabling on-chip learning over second-long timescales,

    C. Frenkel and G. Indiveri, “Reckon: A 28nm sub-mm2 task-agnostic spiking recurrent neural network processor enabling on-chip learning over second-long timescales,” inInternational Solid-State Circuits Conference (ISSCC), vol. 65, pp. 1–3, 2022

    2022

  3. [11]

    Bottom-up and top-down ap- proaches for the design of neuromorphic processing systems: Tradeoffs and synergies between natural and artificial intelligence,

    C. Frenkel, D. Bol, and G. Indiveri, “Bottom-up and top-down ap- proaches for the design of neuromorphic processing systems: Tradeoffs and synergies between natural and artificial intelligence,”Proceedings of the IEEE, vol. 111, no. 6, pp. 623–652, 2023

    2023

  4. [12]

    Efficient video and audio processing with loihi 2,

    S. B. Shrestha, J. Timcheck, P. Frady, L. Campos-Macias, and M. Davies, “Efficient video and audio processing with loihi 2,” in International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 13481–13485, 2024

    2024

  5. [13]

    Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip,

    M. Yao et al., “Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip,”Nature Communications, vol. 15, no. 4464, 2024

    2024

  6. [14]

    Advancing neuromorphic computing with loihi: A survey of results and outlook,

    M. Davies et al., “Advancing neuromorphic computing with loihi: A survey of results and outlook,”Proceedings of the IEEE, vol. 109, no. 5, pp. 911–934, 2021

    2021

  7. [15]

    Seneca: building a fully digital neuromorphic pro- cessor, design trade-offs and challenges,

    G. Tang et al., “Seneca: building a fully digital neuromorphic pro- cessor, design trade-offs and challenges,”Frontiers in Neuroscience, vol. 17, 2023

    2023

  8. [16]

    Towards artificial general intelligence with hybrid tianjic chip architecture,

    J. Pei et al., “Towards artificial general intelligence with hybrid tianjic chip architecture,”Nature, vol. 572, pp. 106–111, 2019

    2019

  9. [17]

    Neuromorphic computing chip with spatiotemporal elasticity for multi-intelligent-tasking robots,

    S. Ma et al., “Neuromorphic computing chip with spatiotemporal elasticity for multi-intelligent-tasking robots,”Science Robotics, vol. 7, no. 67, 2022

    2022

  10. [18]

    The spinnaker project,

    S. Furber, F. Galluppi, S. Temple, and L. Plana, “The spinnaker project,”Proceedings of the IEEE, vol. 102, no. 5, pp. 652–665, 2014

    2014

  11. [19]

    Overview of the spinnaker system architecture,

    S. Furber, D. Lester, L. Plana, J. Garside, E. Painkras, S. Temple, and A. Brown, “Overview of the spinnaker system architecture,”IEEE Transactions on Computers, vol. 62, no. 12, pp. 2454–2467, 2013

    2013

  12. [20]

    Dynamic power management for neuromorphic many-core systems,

    S. H ¨oppner et al., “Dynamic power management for neuromorphic many-core systems,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 66, no. 8, pp. 2973–2986, 2019

    2019

  13. [21]

    SpiNNaker 2: A 10 million core processor system for brain simulation and machine learning,

    C. Mayr, S. Hoeppner, and S. Furber, “SpiNNaker 2: A 10 million core processor system for brain simulation and machine learning,” in Communicating Process Architectures 2017 & 2018, p. 277–280, IOS Press, 2019

    2017

  14. [22]

    Approximate fixed-point elementary function accelerator for the spinnaker-2 neuromorphic chip,

    M. Mikaitis et al., “Approximate fixed-point elementary function accelerator for the spinnaker-2 neuromorphic chip,” in2018 IEEE 25th Symposium on Computer Arithmetic (ARITH), pp. 37–44, 2018

    2018

  15. [23]

    True random number generation from bang-bang adpll jitter,

    F. Neumarker, S. H ¨oppner, A. Dixius, and C. Mayr, “True random number generation from bang-bang adpll jitter,” in2016 IEEE Nordic Circuits and Systems Conference (NORCAS), pp. 1–5, 2016

    2016

  16. [24]

    Stochastic rounding: Algorithms and hardware acceler- ator,

    M. Mikaitis, “Stochastic rounding: Algorithms and hardware acceler- ator,” inIJCNN 2021, pp. 1–6, 2021

    2021

  17. [25]

    A 16-channel fully configurable neural soc with 1.52µw/ch signal acquisition, 2.79µw/ch real-time spike classifier, and 1.79 tops/w deep neural network accelerator in 22 nm fdsoi,

    S. M. A. Zeinolabedin et al., “A 16-channel fully configurable neural soc with 1.52µw/ch signal acquisition, 2.79µw/ch real-time spike classifier, and 1.79 tops/w deep neural network accelerator in 22 nm fdsoi,”IEEE TBioCAS, vol. 16, no. 1, pp. 94–107, 2022

    2022

  18. [26]

    Mapping Deep Neural Networks on SpiNNaker2,

    F. Kelber et al., “Mapping Deep Neural Networks on SpiNNaker2,” inNICE 2020, pp. 1–3, 2020

    2020

  19. [27]

    A tutorial on crc computations,

    T. Ramabadran and S. Gaitonde, “A tutorial on crc computations,” IEEE Micro, vol. 8, no. 4, pp. 62–75, 1988

    1988

  20. [28]

    Aurora 8B/10B Protocol Specifica- tion,

    Advanced Micro Devices, Inc., “Aurora 8B/10B Protocol Specifica- tion,” 2014

    2014

  21. [29]

    UDT: UDP-based data transfer for high-speed wide area networks,

    Y . Gu and R. L. Grossman, “UDT: UDP-based data transfer for high-speed wide area networks,”Computer Networks, vol. 51, no. 7, pp. 1777–1799, 2007

    2007

  22. [30]

    Adaptive body bias aware implementation for ultra-low-voltage designs in 22fdx technology,

    S. H ¨oppner et al., “Adaptive body bias aware implementation for ultra-low-voltage designs in 22fdx technology,”IEEE Transactions on Circuits and Systems II, vol. 67, no. 10, pp. 2159–2163, 2020

    2020

  23. [31]

    How to achieve world-leading energy efficiency using 22fdx with adaptive body biasing on an arm cortex-m4 iot soc,

    S. H ¨oppner et al., “How to achieve world-leading energy efficiency using 22fdx with adaptive body biasing on an arm cortex-m4 iot soc,” inESSDERC 2019, pp. 66–69, 2019

    2019

  24. [32]

    py-spinnaker2,

    B. V oggingeret al., “py-spinnaker2,” 2024. https://zenodo.org/doi/10. 5281/zenodo.10202109

    2024

  25. [33]

    Pynn: a common interface for neuronal network simulators,

    A. P. Davison, D. Br ¨uderle, J. M. Eppler, J. Kremkow, E. Muller, D. Pecevski, L. Perrinet, and P. Yger, “Pynn: a common interface for neuronal network simulators,”Frontiers in neuroinformatics, vol. 2, p. 388, 2009

    2009

  26. [34]

    OctopuScheduler: On-Chip DNN Scheduling on the SpiNNaker2 Neuromorphic MPSoC,

    T. Langer, M. Jobst, C. Liu, F. Kelber, B. V ogginger, and C. Mayr, “OctopuScheduler: On-Chip DNN Scheduling on the SpiNNaker2 Neuromorphic MPSoC,” in2025 Neuro Inspired Computational El- ements (NICE), pp. 1–10, Mar. 2025

    2025

  27. [35]

    An end-to-end dnn inference framework for the spinnaker2 neuro- morphic mpsoc,

    M. Jobst, T. Langer, C. Liu, M. Alici, H. A. Gonzalez, and C. Mayr, “An end-to-end dnn inference framework for the spinnaker2 neuro- morphic mpsoc,” inICONS 2025, 2025

    2025

  28. [36]

    Coremark benchmark,

    EEMBC, “Coremark benchmark,” 2025. Accessed: 2025-01-15

    2025

  29. [37]

    Cmsis-nn: Efficient neural network kernels for arm cortex-m cpus,

    L. Lai, N. Suda, and V . Chandra, “Cmsis-nn: Efficient neural network kernels for arm cortex-m cpus,”arXiv:1801.06601, 2018

    2018 arXiv

  30. [38]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” Dec. 2015. arXiv:1512.03385 [cs]

    2015 arXiv

  31. [39]

    OPT: Open Pre-trained Transformer Language Models,

    S. Zhang et al., “OPT: Open Pre-trained Transformer Language Models,” 2022. arXiv:2205.01068 [cs]

    2022 arXiv

  32. [40]

    NVIDIA Jetson Orin Nano Super Developer Kit. Datasheet,

    NVIDIA Corporation, “NVIDIA Jetson Orin Nano Super Developer Kit. Datasheet,” 2024

    2024

  33. [41]

    11.4 IBM NorthPole: An Architecture for Neural Network Inference with a 12nm Chip,

    A. S. Cassidy et al., “11.4 IBM NorthPole: An Architecture for Neural Network Inference with a 12nm Chip,” in2024 IEEE International Solid-State Circuits Conference (ISSCC), vol. 67, pp. 214–215, 2024

    2024

  34. [42]

    Neural inference at the frontier of energy, space, and time,

    D. S. Modha et al., “Neural inference at the frontier of energy, space, and time,”Science, vol. 382, pp. 329–335, Oct. 2023

    2023

  35. [43]

    sPyNNaker: A software package for running PyNN simulations on SpiNNaker,

    O. Rhodes et al., “sPyNNaker: A software package for running PyNN simulations on SpiNNaker,”Frontiers in Neuroscience, vol. 12, 2018

    2018

  36. [44]

    Efficient Deployment of Spiking Neural Networks on SpiNNaker2 for DVS Gesture Recognition Using Neuromorphic Intermediate Rep- resentation,

    S. Arfa, B. V ogginger, C. Liu, J. Partzsch, M. Sch ¨one, and C. Mayr, “Efficient Deployment of Spiking Neural Networks on SpiNNaker2 for DVS Gesture Recognition Using Neuromorphic Intermediate Rep- resentation,” inNICE 2025, pp. 1–8, 2025

    2025

  37. [45]

    A low power, fully event-based gesture recognition system,

    A. Amiret al., “A low power, fully event-based gesture recognition system,” inIEEE conference on computer vision and pattern recogni- tion, pp. 7243–7252, 2017

    2017

  38. [46]

    Neuromorphic intermediate representation: A unified instruction set for interoperable brain-inspired computing,

    J. E. Pedersenet al., “Neuromorphic intermediate representation: A unified instruction set for interoperable brain-inspired computing,” Nature Communications, vol. 15, p. 8122, Sept. 2024

    2024

  39. [47]

    Spinnaker2: A large-scale neuromorphic sys- tem for event-based and asynchronous machine learning,

    H. A. Gonzalezet al., “Spinnaker2: A large-scale neuromorphic sys- tem for event-based and asynchronous machine learning,” inMLNPCP Workshop at NeurIPS 2023, 2023

    2023

  40. [48]

    Efficient recurrent architectures through activity sparsity and sparse back-propagation through time,

    A. Subramoney, K. K. Nazeer, M. Sch ¨one, C. Mayr, and D. Kappel, “Efficient recurrent architectures through activity sparsity and sparse back-propagation through time,” inICLR, 2023

    2023

  41. [49]

    Language modeling on a SpiNNaker2 neuromor- phic chip,

    K. K. Nazeeret al., “Language modeling on a SpiNNaker2 neuromor- phic chip,” inIEEE AICAS, pp. 492–496, 2024

    2024

  42. [50]

    Event-based backpropagation on the neuromorphic plat- form spinnaker2,

    G. B ´ena, T. Wunderlich, M. Akl, B. V ogginger, C. Mayr, and H. A. Gonzalez, “Event-based backpropagation on the neuromorphic plat- form spinnaker2,” in2025 Neuro Inspired Computational Elements (NICE), pp. 1–10, IEEE, 2025

    2025

  43. [51]

    E-prop on SpiN- Naker 2: Exploring online learning in spiking RNNs on neuromorphic hardware,

    A. Rostami, B. V ogginger, Y . Yan, and C. G. Mayr, “E-prop on SpiN- Naker 2: Exploring online learning in spiking RNNs on neuromorphic hardware,”Frontiers in Neuroscience, vol. 16, 2022

    2022

  44. [52]

    A solution to the learning dilemma for recurrent networks of spiking neurons,

    G. Bellecet al., “A solution to the learning dilemma for recurrent networks of spiking neurons,”Nature Communications, vol. 11, no. 1, p. 3625, 2020

    2020

  45. [53]

    ON-OFF neuromorphic ISING machines using fowler- nordheim annealers,

    Z. Chenet al., “ON-OFF neuromorphic ISING machines using fowler- nordheim annealers,”Nature Communications, vol. 16, 2025

    2025

  46. [54]

    Deploying Machine Learning Models to Ahead-of-Time Runtime on Edge Using MicroTVM,

    C. Liu, M. Jobst, L. Guo, X. Shi, J. Partzsch, and C. Mayr, “Deploying Machine Learning Models to Ahead-of-Time Runtime on Edge Using MicroTVM,” inCODAI Workshop, p. 37–40, 2024

    2024

  47. [55]

    Neuromorphic computing at scale,

    D. Kudithipudi, C. Schuman, C. M. Vineyard, T. Pandit, C. Merkel, R. Kubendran, J. B. Aimone, G. Orchard, C. Mayr, R. Benosman, et al., “Neuromorphic computing at scale,”Nature, vol. 637, no. 8047, pp. 801–812, 2025

    2025

  48. [56]

    Efficient snn multi-cores mac array acceleration on spinnaker 2,

    J. Huang, F. Kelber, B. V ogginger, C. Liu, F. Kreutz, P. Gerhards, D. Scholz, K. Knobloch, and C. G. Mayr, “Efficient snn multi-cores mac array acceleration on spinnaker 2,”Frontiers in Neuroscience, vol. 17, 2023. 14 VOLUME , <Society logo(s) and publication title will appear here.>

    2023

  49. [57]

    Nestml: a generic modeling language and code generation tool for the simulation of spiking neural networks with advanced plasticity rules,

    C. Linssen, P. N. Babu, J. M. Eppler, L. Koll, B. Rumpe, and A. Mor- rison, “Nestml: a generic modeling language and code generation tool for the simulation of spiking neural networks with advanced plasticity rules,”Frontiers in Neuroinformatics, vol. V olume 19 - 2025, 2025

    2025

  50. [58]

    Real-time cortical simulation on neu- romorphic hardware,

    O. Rhodes, L. Peres, A. G. Rowley, A. Gait, L. A. Plana, C. Bren- ninkmeijer, and S. B. Furber, “Real-time cortical simulation on neu- romorphic hardware,”Philosophical Transactions of the Royal Society A, vol. 378, no. 2164, p. 20190160, 2020

    2020

  51. [59]

    Synapse-centric mapping of cortical models to the spinnaker neuromorphic architecture,

    J. C. Knight and S. B. Furber, “Synapse-centric mapping of cortical models to the spinnaker neuromorphic architecture,”Frontiers in Neuroscience, vol. V olume 10 - 2016, 2016

    2016

  52. [60]

    Parallelization of neural processing on neuromorphic hardware,

    L. Peres and O. Rhodes, “Parallelization of neural processing on neuromorphic hardware,”Frontiers in Neuroscience, vol. V olume 16 - 2022, 2022

    2022

  53. [61]

    E-prop on spin- naker 2: Exploring online learning in spiking rnns on neuromorphic hardware,

    A. Rostami, B. V ogginger, Y . Yan, and C. G. Mayr, “E-prop on spin- naker 2: Exploring online learning in spiking rnns on neuromorphic hardware,”Frontiers in Neuroscience, vol. 16, 2022

    2022

  54. [62]

    Speech Commands: A Dataset for Limited-V ocabulary Speech Recognition,

    P. Warden, “Speech Commands: A Dataset for Limited-V ocabulary Speech Recognition,”arXiv:1804.03209 [cs], Apr. 2018. VOLUME , 15 : VI. Supplementary Material for the paper ”The SpiNNaker2 chip: a many-core platform for flexible and scalable brain-inspired computing” A. Interna...

    2018 arXiv

  55. [63]

    TCAM Lookup Stage: Performs associative matching of packet address field against routing entries

  56. [64]

    Priority Encoding Stage: Resolves multiple matches and selects the highest-priority routing entry

  57. [65]

    Link-Destination Resolution Stage: Determines the set of 7 off-chip links

  58. [66]

    Core-Destination Lookup Stage: Determines internal 152 core destinations. To improve energy efficiency, the fourth stage can be dynamically disabled when no internal core destinations are present, reducing unnecessary lookup activity and lowering dynamic power consumption. c: ...

  59. [67]

    Destination Field (8 bits), distinguishing between the seven off-chip links, the monitor core, other cores and dropped packets

  60. [68]

    Source Field (2 bits), differentiating between off-chip packets and on-chip packets originating from process- ing elements (PEs)

  61. [69]

    Payload Size Field (4 bits), encoding different packet lengths: packets without payload, packets with 32-bit payload, 64-bit payload or 128-bit payload

  62. [70]

    Routing Algorithm Field (2 bits), indicating whether default routing was applied or a non-default routing scheme was used

  63. [71]

    Packet Type Field (3 bits), classifying packets as Near- est Neighbour (NN), Core-to-Core (C2C) or Multicast (MC) packets. In addition to packet classification counters, the SpiN- Naker2 router integrates several dedicated hardware profiling counters to support performance ana...

  64. [72]

    Router Cycle Counter: Measures the number of clock cycles consumed during a specified router operation

  65. [73]

    Busy Cycle Counter: Counts router wait cycles, i.e., cycles during which backpressure occurs at any of the six router input ports

  66. [74]

    Zero-Wait Packet Counter: Records the number of packets that traverse one of the MC, C2C, or NN routing engine outputs without incurring any delay

  67. [75]

    Iterative Drop Counter: Counts packets that are dropped more than once during routing

  68. [76]

    Error Packet Counter: Records packets associated with routing errors, including time-phase violations and unroutable packet errors

  69. [77]

    Total Reinserted Packet Counter: Counts the number of packets that are reinserted into the routing fabric by hardware mechanisms

  70. [78]

    The associated water-level reg- ister records the maximum occupancy reached in this buffer

    Dropped packet Water-Level Counter: Tracks the num- ber of SpiNNaker packets currently stored in the dropped-packet buffer. The associated water-level reg- ister records the maximum occupancy reached in this buffer. C. Diagnostic Counters for Event Links Each short-range and l...

  71. [79]

    Error in CRC of received packet

  72. [80]

    Correct sequence ID in received packet

  73. [81]

    Link partner requested re-send

  74. [82]

    Wrong header in received packet Each counter is 8 bit wide. D. Event Handler and Spike FIFO As mentioned in Section III.B.1 and Fig. 3, each PE has an event handler to process incoming SpiNNaker packets without interrupting the processor. The event handler has two filters and ...

  75. [83]

    Synchronization The synchronization of PEs for SNNs works as follows:

  76. [84]

    Then, the PE goes into WFI (wait for interrupt) state

    For each PE, the ARM program is loaded and started to call setup functions according to configuration data stored in SRAM. Then, the PE goes into WFI (wait for interrupt) state

  77. [85]

    Then, the PEs start the timer to trigger a timer interrupt everyNreference clock cycles, for example N= 1000for a1 mstime step considering the1 MHz reference clock, cf

    A chip-global interrupt wakes up all PEs simultane- ously. Then, the PEs start the timer to trigger a timer interrupt everyNreference clock cycles, for example N= 1000for a1 mstime step considering the1 MHz reference clock, cf. Section III.B.1. Again, the PEs go into WFI state

  78. [86]

    Then, the program starts to process these spikes, performs synapse and neuron updates, sends out spikes, and records spikes and voltages (if enabled)

    The timer interrupt wakes up all PEs: The neuron cores first read the current spike FIFO’s read and write pointer position to know which spikes were received in the previous time step. Then, the program starts to process these spikes, performs synapse and neuron updates, sends...

  79. [87]

    9: We used the high-level library for SNNpy-spinnaker2[32] for the characterization

    SNN characterization Details for the study in Fig. 9: We used the high-level library for SNNpy-spinnaker2[32] for the characterization. Two different neuron models are used:lif_neuron– leaky integrate and fire (LIF) neuron with sparse incoming connections (individual synapses ...

  80. [88]

    SNN application For the Deep SNN in section IV .D we used the following partitioning and placement strategy of populations to SpiN- Naker2 PEs: Inpy-spinnaker2, we manually defined the maximum number of neurons per core (PE) for each population as given in Tab. 7. The software...

  81. [89]

    A similar capacity study like the one shown in Fig

    Comparison to SpiNNaker While it would be desirable to provide quantitative numbers on the advancements of SpiNNaker2 over SpiNNaker, a fair comparison of the neuron and synapse processing capacity is currently not possible due to different software stacks. A similar capacity ...

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.