GATMesh: Clock Mesh Timing Analysis using Graph Neural Networks

Matthew Guthaus; Muhammad Hadir Khan

REVIEW 4 major objections 6 minor 27 references

GATMesh: Clock Mesh Timing Analysis using Graph Neural Networks

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A graph neural network trained on synthetic meshes predicts clock mesh delay within about 5 ps of SPICE at 47,000x speedup.

desk verdict A thoughtful first GNN for clock mesh timing, with an honest evaluation that overclaims generalization beyond its training area range. read the letter →

arxiv 2507.05681 v1 pith:FHGLGOJX submitted 2025-07-08 cs.AR cs.AIcs.LG

Muhammad Hadir Khan , Matthew Guthaus This is my paper

classification cs.ARcs.AIcs.LG

keywords clockmeshgraphneuralnetworkdelaypredictionslewjumpingknowledgeVLSItiminganalysisskewSPICE

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

This paper argues that clock mesh timing—the delay and slew at every clock sink—can be predicted accurately and almost instantly by a graph neural network trained once on SPICE-simulated synthetic meshes. If true, it removes the main obstacle to using clock meshes in automated design flows: SPICE-accurate analysis is too slow, while simplified first-order models miss slew and buffer contention effects. The model, GATMesh, reports average delay error of 5.27 ps and average slew error of 5.98 ps on unseen open-source benchmarks, measured against SPICE, with a runtime speedup of roughly 47,000x over multi-threaded SPICE. The paper also shows that design-inspired auxiliary graph edges and learnable Jumping Knowledge connections each contribute to accuracy.

What carries the argument

The clock mesh is encoded as an undirected graph whose nodes are the physical objects of the mesh (buffers, taps, wires, and sinks) plus auxiliary connection nodes; message passing runs along mesh edges and auxiliary edges. Each sink carries two regression targets, output delay and output slew. Two hand-designed auxiliary edge types shortcut the graph: each sink connects to its two nearest driving buffers in path-resistance terms, modeling multi-driver sink driving, and each buffer connects to its four nearest buffers, modeling buffer contention from skewed input arrivals. Node features include capacitance, resistance, total and minimum path resistance from buffers, regional capacitance, layout coordinates, and buffer input delay and slew. The model stacks eight graph attention convolution layers with four attention heads and aggregates layer outputs with Jumping Knowledge max-pooling, which selects the most informative receptive field per feature coordinate. A masked mean-squared-error loss on sink nodes trains the model against SPICE labels.

What would settle it

Measure the average delay mean-absolute error on a set of synthetic meshes generated with areas about five and ten times the largest training area, using the paper's own mesh-generation pipeline; if the average error exceeds the reported worst-case value of about 16 ps, the claimed few-picosecond accuracy is confined to the training distribution and does not generalize to larger designs.

Watch

Extended reading notes

Core claim

The paper's discovery is that a single graph neural network, trained once on synthetic meshes generated from the area, sink count, and sink-density statistics of open-source designs, can inductively predict the delay and slew at every clock sink of previously unseen real designs with average absolute errors of 5.27 ps and 5.98 ps respectively, measured against SPICE. The same model is about 47,000x faster than multi-threaded SPICE on the tested designs and about 1,700x faster than the first-order model, while the first-order model misses slew entirely and has an average delay error of about 200 ps. The ablation demonstrates that both the static auxiliary connections, which model sink driving and buffer contention, and the learned Jumping Knowledge connections contribute to accuracy, with auxiliary connections providing most of the delay improvement.

Load-bearing premise

The synthetic meshes used for training, whose sizes and sink densities are drawn from statistics of small open-source designs, are representative enough of unseen designs that a model trained on them predicts delays and slews within a few picoseconds of SPICE.

Editorial extensions

If this is right

Clock mesh delay and slew can be estimated to within a few picoseconds of SPICE without simulating the mesh, making repeated timing checks during mesh design practical.
Because slew is predicted alongside delay, setup and hold checks can directly account for clock slew effects, which simplified mesh models omit.
The runtime advantage of tens of thousands of times over multi-threaded SPICE means clock mesh timing analysis can run inside optimization loops at design time rather than only as final verification.
The ablation shows that removing either the static auxiliary edges or the Jumping Knowledge connections degrades prediction accuracy, so future mesh-learning models should include both mechanisms.

Reading between the lines

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

An implication the paper leaves implicit is that the same graph encoding could apply to other multi-driver, reconvergent RC networks such as power grids or resonant clock distribution, where delay and slew also depend on simultaneous drivers and contention.
A testable extension would be to retrain or fine-tune on out-of-distribution designs; the paper's own worst-case numbers, about 16 ps delay and 13.5 ps slew on a design far larger than any training mesh, suggest error grows with distance from the training distribution, but the paper does not quantify that growth.
The speedup over SPICE suggests the model could serve as a differentiable surrogate inside mesh sizing and buffer insertion optimization, replacing simplified delay-proxy objectives with learned SPICE-accurate ones; the paper does not demonstrate this optimization loop.
Because the loss is computed only on sink nodes, the internal representations on buffer and wire nodes are free to encode physical quantities; probing these embeddings might reveal whether the model learns interpretable notions of driving resistance or local capacitance.

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.

Desk Editor's Note

read the letter

GATMesh is the first GNN approach for clock mesh timing analysis I've seen, and the domain adaptation is genuinely thoughtful. The authors model the mesh as an RC graph and add two auxiliary connection sets: one from each sink to its two nearest driving buffers, and one between each buffer and its four nearest buffers to capture contention. Combined with features like region capacitance and total/min path resistance, this is a sensible encoding of the physical problem. The ablation shows the auxiliary connections do most of the work, which fits the intuition that message passing to nearby drivers matters more than deeper learned aggregation.

The evaluation is more honest than most in this area. They compare against SPICE and a first-order model, report per-design errors, and explicitly point out that the worst case (swerv wrapper, 16.16 ps delay MAE, 13.50 ps slew MAE) is far outside the training distribution. That self-criticism is real and matches the stress-test note: excluding swerv wrapper drops the average delay error from 5.27 ps to 3.72 ps. So the headline accuracy is sensitive to one out-of-distribution point, and the paper admits it. The 'unseen benchmarks' claim should be read as 'unseen designs within the same OpenROAD/Nangate45 flow and area range.' That is a narrower claim than the abstract implies.

The main soft spots: no code or data release, so the results are not independently reproducible; only eight test designs; no error bars or multiple training runs; and the runtime comparison mixes GPU inference with CPU simulation. None of these are fatal to the core idea, but they limit how much I trust the generalization numbers.

This is not a paradigm shift, but it is a useful step forward for a real problem. I would send this to peer review at a place like DAC or ICCAD, with a request for code/data, out-of-distribution stress tests, and uncertainty estimates. I would not desk-reject it. For someone working on clock distribution or GNN-based timing, it is a citable reference, even though I would not build directly on it until the artifact is available.

Referee Report

4 major / 6 minor

Summary. The paper proposes GATMesh, a graph neural network for clock-mesh timing analysis. The clock mesh is modeled as an undirected graph with buffer, tap, wire, sink, and auxiliary nodes; node features include input delay/slew, capacitance, resistance, region capacitance, and layout coordinates. The model uses eight GATConv layers with four attention heads and Jumping Knowledge connections, trained with MSE loss against SPICE-computed sink delay and slew labels on 900 synthetic meshes. On eight OpenROAD designs unseen during training, it reports average delay MAE of 5.27 ps, average slew MAE of 5.98 ps, and an average 47,146x speedup over Ngspice. An ablation study compares the full model with variants lacking auxiliary connections and/or JK connections.

Significance. If the accuracy claims hold beyond the reported benchmarks, GATMesh would be a useful step toward fast, SPICE-comparable clock-mesh analysis in EDA flows, and it is one of the first GNN treatments of multi-driver, reconvergent clock meshes. The paper's strengths include a physically motivated graph construction with auxiliary connections for sink-driving and buffer contention, direct held-out evaluation against SPICE in Table IV, a concrete runtime comparison in Table V, and an ablation that separates the contributions of the auxiliary and JK components. However, the paper provides no code, model, or trained weights, and the central generalization claim is supported by only eight test designs with no uncertainty quantification and one clearly out-of-distribution worst case. The reported few-picosecond accuracy is real for the eight listed designs but is not established for designs outside the training area range.

major comments (4)

[§V.B, Table IV] The headline average delay MAE of 5.27 ps is dominated by an out-of-distribution test design. Training areas are capped at 220,000 µm² (Table II), yet swerv wrapper has area 1,056,664 µm² and delay MAE 16.16 ps, and §V.B itself concedes this is because it is "significantly larger area compared to our training statistics." Removing this single design lowers the average delay MAE to 3.72 ps. Since real high-performance meshes can exceed the 220k µm² cap, the claim of few-picosecond accuracy on unseen benchmarks is only supported within the training area distribution; the paper should either add large-area training examples or report accuracy separately by area range.
[§V.A, §V.B] The evaluation has no uncertainty quantification. Eight test designs each provide a single SPICE-vs-model error, the model is trained once with no repeated seeds, and no confidence intervals are reported. The average MAE is sensitive to individual designs, as shown by the swerv wrapper case. The authors should report standard deviations over repeated training runs or per-design confidence intervals so the reader can assess whether the 5.27 ps and 5.98 ps averages are stable.
[§V.C, Table V] The runtime speedup comparison is inference-only on an NVIDIA RTX 4090, while the abstract's "47146x over multi-threaded SPICE simulation" does not account for the roughly two-hour training time or the SPICE label generation needed to create the training set. A fair statement should either amortize these offline costs or explicitly scope the claim to inference. The comparison is also across different hardware platforms (GPU inference versus CPU SPICE), so the speedup number is platform-specific and should be presented as such.
[§V.D, Figure 5] The ablation study is performed on the synthetic training dataset, not on the eight open-source designs of Table IV. Consequently, the ablation does not directly establish that the auxiliary connections and JK connections improve generalization to unseen realistic designs. Since the text states that auxiliary connections alone give slightly better delay MAE than the full GATMesh on the synthetic test set, the contribution of JK connections to the real-design accuracy needs direct evaluation on the open-source benchmarks.

minor comments (6)

[§III title] The section heading appears as "GATM ESH" instead of "GATMesh".
[§III.E] The phrase "hf inal" should be written as "h_final" or "h_final"; there also appears to be a missing closing brace in the surrounding text.
[§V.A] MSE losses are reported as "23.87 ps", "31.83 ps", and "33.02 ps", but MSE has units of ps²; please correct the units.
[Table II] For the "Large" row, the density is a fixed value 0.021, while the "Small" and "Medium" rows give ranges; it would be clearer to state why the large designs use a single density value.
[Availability] No code, trained model, or dataset is provided, which limits reproducibility of the reported accuracy and speedup numbers.
[Table IV] It is unclear how the open-source test designs' clock meshes were generated (mesh size, buffering type, and top-level tree construction) and whether they match the synthetic training generation methodology; please describe the test design generation in the text or a footnote.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GATMesh accuracy is benchmarked against independent Ngspice SPICE labels.

full rationale

GATMesh is a supervised regression model whose sink delay/slew labels are produced by Ngspice SPICE simulation (Section IV.B: 'the target labels which are computed by running SPICE simulation on the clock meshes'), and whose reported accuracy is measured against the same external SPICE reference on held-out OpenROAD designs (Section V.B, Table IV). No equation in the paper defines a prediction as a fitted input; the learned GNN maps structural and physical features (Table I) to SPICE-derived labels, and the test designs are not used in training or hyperparameter selection. The self-citations to [6], [9], and [10] are used only to motivate mesh generation and buffer-contention auxiliary connections; they do not supply the accuracy result, which is benchmarked against Ngspice. The out-of-distribution swerv wrapper issue noted in Section V.B is a generalization limitation, not circularity. Therefore no circular step can be identified.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim depends on: (1) Ngspice as ground truth, (2) DME and Elmore-based top-level tree delays as adequate input features, (3) synthetic meshes matching the test distribution, and (4) hand-chosen auxiliary connection counts. The model is a supervised surrogate, so the learned weights are fitted parameters in a standard sense; the hand-chosen graph construction choices are the main ad hoc elements.

free parameters (6)

number of nearest buffers per sink for auxiliary connections = 2
Section III.A: 'in practice, we have found that considering the two nearest buffers for every sink adds appropriate insight into the neighborhood of buffers driving a sink.' This affects graph connectivity and the reported accuracy; it is a hand-chosen heuristic.
number of nearest buffers per buffer for buffer contention connections = 4
Section III.B: 'in practice, we have found that adding the four nearest buffers to every buffer adds appropriate insight into the neighborhood of buffers.' This choice affects message passing between buffers and is justified by design intuition, not by a formal argument.
resistance limit for the region cap feature = resistance of a single mesh wire
Section III.C: 'We chose the resistance limit as the resistance of a single mesh wire so that it will consider all of the nearby sinks.' This defines a derived feature used by the model.
number of GATConv layers = 8
Section IV.C: 'Through empirical evaluations, we found better performance by having a deep model architecture and therefore used eight convolution layers.' Tuned on validation data.
number of attention heads = 4
Section IV.C: 'We observed that four attention heads performed well, likely due to the four orthogonal directions at each mesh point.' Tuned on validation data.
JK aggregation type = max
Section III.E: of concatenation, max pooling, and Bi-directional LSTM, the authors chose max pooling, citing better performance in deep GNNs. This is an architecture choice affecting the learned representation.

assumptions (5)

domain assumption Ngspice simulation is the golden reference for sink delay and slew
The paper trains and validates against Ngspice (Sections IV.A and V.B); all accuracy claims are relative to this simulator. If Ngspice differs from commercial SPICE or silicon, the reported errors are relative to that baseline.
domain assumption Top-level tree input delay and slew computed with DME and Elmore models are adequate input features
Section IV.A: 'The top-level trees are zero-skew according to the Elmore model but may have some skew due to mismatch between SPICE and Elmore delay models.' The model uses these approximate values as features for mesh buffers; the learned mapping must absorb this mismatch.
domain assumption Synthetic training meshes are representative of the test open-source meshes
Section IV.A generates synthetic designs from statistics of OpenROAD designs; the test set consists of distinct OpenROAD designs. The paper itself notes the worst case 'swerv wrapper' has area outside the training statistics (Section V.B), indicating the assumption can fail.
ad hoc to paper The two-nearest-buffer and four-nearest-buffer auxiliary connections suffice to capture multi-driver and buffer-contention effects
Sections III.A and III.B justify these choices by design intuition ('makes sense because each sink is on a mesh wire...'), not by a formal argument; the ablation shows auxiliary connections help, but the exact counts are a modeling choice.
standard math GAT and JK connection behavior follows the cited prior work
The model relies on the Graph Attention Network mechanism of [18] and Jumping Knowledge connections of [14]; no proof or re-derivation is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GATMesh: Clock Mesh Timing Analysis using Graph Neural Networks." pith.science (2026). https://pith.science/paper/FHGLGOJX

@misc{pith2026250705681,
  author       = {Pith},
  title        = {Pith review of: GATMesh: Clock Mesh Timing Analysis using Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FHGLGOJX}},
  note         = {Machine review of arXiv:2507.05681}
}

read the original abstract

Clock meshes are essential in high-performance VLSI systems for minimizing skew and handling PVT variations, but analyzing them is difficult due to reconvergent paths, multi-source driving, and input mesh buffer skew. SPICE simulations are accurate but slow; yet simplified models miss key effects like slew and input skew. We propose GATMesh, a Graph Neural Network (GNN)-based framework that models the clock mesh as a graph with augmented structural and physical features. Trained on SPICE data, GATMesh achieves high accuracy with average delay error of 5.27ps on unseen benchmarks, while achieving speed-ups of 47146x over multi-threaded SPICE simulation.

Figures

Figures reproduced from arXiv: 2507.05681 by the authors.

**Figure 2.** Left: A single mesh wire, its buffers and its graph representation. Right: Graph representation after adding buffer contention (X0) and sink driving (X1 − X4) auxiliary connections. through the intermediate T0 and T1 nodes and the divided W1 and W3 node resulting from the split mesh wire. Any GNN model would require an increasingly deeper number of layers to allow message passing through deep graphs, yet so many dee… view at source ↗

**Figure 3.** Clock signal input slew has a significant impact on setup and hold [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Model architecture overview. As an example three GATConv layers are shown with Jumping Knowledge connections (shown in red). Masking is [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: The addition of the auxiliary connections offers the most benefit for [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 21 canonical work pages

[1]

Sizing of clock distribution networks for high performance CPU chips,

M. P. Desai, R. Cvijetic, and J. Jensen, “Sizing of clock distribution networks for high performance CPU chips,” in Design Automation Conference (DAC), pp. 389–394, 1996

work page 1996
[2]

The clock distribution of the POWER4 mi- croprocessor,

P. Restle, C. Carter, J. Eckhardt, B. Krauter, B. McCredie, K. Jenkins, A. Weger, and A. Mule, “The clock distribution of the POWER4 mi- croprocessor,” in International Solid-State Circuits Conference (ISSCC), pp. 144–145, 2002

work page 2002
[3]

A 5GHz duty-cycle correcting clock distribution network for the POWER6 microprocessor,

M. G.-R. Thomson, P. J. Restle, and N. K. James, “A 5GHz duty-cycle correcting clock distribution network for the POWER6 microprocessor,” in International Solid-State Circuits Conference (ISSCC) , pp. 1522– 1529, 2006

work page 2006
[4]

Resonant clock design for a power-efficient high-volume x86–64 microprocessor,

V . Sathe, S. Arekapudi, C. Ouyang, M. Papaefthymiou, A. Ishii, and S. Naffziger, “Resonant clock design for a power-efficient high-volume x86–64 microprocessor,” in2012 IEEE International Solid-State Circuits Conference, pp. 68–70, 2012

work page 2012
[5]

Design of the two-core x86-64 amd “bulldozer

H. McIntyre, S. Arekapudi, E. Busta, T. Fischer, M. Golden, A. Horiuchi, T. Meneghini, S. Naffziger, and J. Vinh, “Design of the two-core x86-64 amd “bulldozer” module in 32 nm soi cmos,” IEEE Journal of Solid- State Circuits, vol. 47, no. 1, pp. 164–176, 2012

work page 2012
[6]

High- performance clock mesh optimization,

M. Guthaus, X. Hu, G. Wilke, G. Flache, and R. Reis, “High- performance clock mesh optimization,” ACM Transactions on Design Automation of Electronic Systems (TODAES) , 2012

work page 2012
[7]

Meshworks: an efficient framework for planning, synthesis and optimization of clock mesh networks,

A. Rajaram and D. Z. Pan, “Meshworks: an efficient framework for planning, synthesis and optimization of clock mesh networks,” in Asia and South Pacific Design Automation Conference (ASP-DAC) , pp. 250– 257, 2008

work page 2008
[8]

Clock mesh synthesis method using the earth mover’s distance under transformations,

Y . Teng and B. Taskin, “Clock mesh synthesis method using the earth mover’s distance under transformations,” in 2012 IEEE 30th International Conference on Computer Design (ICCD) , pp. 121–126, 2012

work page 2012

Show all 27 references

[9]

A novel scheme to reduce short-circuit power in mesh-based clock architectures,

G. Wilke, R. Fonseca, C. Mezzomo, and R. Reis, “A novel scheme to reduce short-circuit power in mesh-based clock architectures,” in Symposium on Integrated Circuits and System Design (SBCCI), pp. 117– 122, 2008

2008
[10]

Non-uniform clock mesh optimization with linear programming buffer insertion,

M. R. Guthaus, G. Wilke, and R. Reis, “Non-uniform clock mesh optimization with linear programming buffer insertion,” in Design Au- tomation Conference (DAC), 2010

2010
[11]

Timing- driven variation-aware nonuniform clock mesh synthesis,

A. Abdelhadi, R. Ginosar, A. Kolodny, and E. G. Friedman, “Timing- driven variation-aware nonuniform clock mesh synthesis,” in Great lakes Symposium on VLSI (GLSVLSI) , pp. 15–20, 2010

2010
[12]

A survey on oversmooth- ing in graph neural networks,

T. K. Rusch, M. M. Bronstein, and S. Mishra, “A survey on oversmooth- ing in graph neural networks,” arXiv:2303.10993, 2023

2023 arXiv
[13]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proceedings of the 31st International Conference on Neural Information Processing Systems , NIPS’17, (Red Hook, NY , USA), p. 6000–6010, Curran...

2017
[14]

Representation learning on graphs with jumping knowledge networks,

K. Xu, C. Li, Y . Tian, T. Sonobe, K. ichi Kawarabayashi, and S. Jegelka, “Representation learning on graphs with jumping knowledge networks,” arXiv:1806.03536, 2018

2018 arXiv
[15]

Bufformer: A generative ml framework for scalable buffering,

R. Liang, S. Nath, A. Rajaram, J. Hu, and H. Ren, “Bufformer: A generative ml framework for scalable buffering,” in Proceedings of the 28th Asia and South Pacific Design Automation Conference , ASPDAC ’23, (New York, NY , USA), p. 264–270, Association for Computing Machinery, 2023

2023
[16]

Fast and accurate wire timing estimation based on graph learning,

Y . Ye, T. Chen, Y . Gao, H. Yan, B. Yu, and L. Shi, “Fast and accurate wire timing estimation based on graph learning,” Design, Automation & Test in Europe Conference (DATE) , 2023

2023
[17]

Fast and accurate wire timing estimation on tree and non-tree net structures,

H.-H. Cheng, I. H.-R. Jiang, and O. Ou, “Fast and accurate wire timing estimation on tree and non-tree net structures,” Design Automation Conference (DAC), 2020

2020
[18]

Graph Attention Networks,

P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph Attention Networks,” arXiv:1710.10903, 2018

2018 arXiv
[19]

The OpenROAD project: Unleashing hardware innovation,

A. B. Kahng and T. Spyrou, “The OpenROAD project: Unleashing hardware innovation,” in Proc. GOMAC, 2021

2021
[20]

Zero-skew clock routing trees with minimum wirelength,

K. Boese and A. Kahng, “Zero-skew clock routing trees with minimum wirelength,” in ASIC Conf., pp. 1.1.1–1.1.5, 1992

1992
[21]

Zero skew clock net routing,

T.-H. Chao, Y .-C. Hsu, and J. Ho, “Zero skew clock net routing,” in Design Automation Conference (DAC) , pp. 518–523, 1992

1992
[22]

Zero skew clock-tree optimiza- tion with buffer insertion/sizing and wire sizing,

J.-L. Tsai, T.-H. Chen, and C. C. Chen, “Zero skew clock-tree optimiza- tion with buffer insertion/sizing and wire sizing,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , vol. 23, no. 4, pp. 565–573, 2004

2004
[23]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” 2017

2017
[24]

Fast and accurate deep network learning by exponential linear units (elus),

D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (elus),” 2016

2016
[25]

PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation,

J. Ansel et al., “PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation,” in 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS ’24), ACM, Apr. 2024

2024
[26]

Fast graph representation learning with PyTorch Geometric,

M. Fey and J. E. Lenssen, “Fast graph representation learning with PyTorch Geometric,” in ICLR Workshop on Representation Learning on Graphs and Manifolds , 2019

2019
[27]

ngspice – open source circuit simulator, version 42

The NGSpice Project, “ngspice – open source circuit simulator, version 42.” https://ngspice.sourceforge.io/, 2024. Accessed: 2025-04-12

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Sizing of clock distribution networks for high performance CPU chips,

M. P. Desai, R. Cvijetic, and J. Jensen, “Sizing of clock distribution networks for high performance CPU chips,” in Design Automation Conference (DAC), pp. 389–394, 1996

work page 1996

[2] [2]

The clock distribution of the POWER4 mi- croprocessor,

P. Restle, C. Carter, J. Eckhardt, B. Krauter, B. McCredie, K. Jenkins, A. Weger, and A. Mule, “The clock distribution of the POWER4 mi- croprocessor,” in International Solid-State Circuits Conference (ISSCC), pp. 144–145, 2002

work page 2002

[3] [3]

A 5GHz duty-cycle correcting clock distribution network for the POWER6 microprocessor,

M. G.-R. Thomson, P. J. Restle, and N. K. James, “A 5GHz duty-cycle correcting clock distribution network for the POWER6 microprocessor,” in International Solid-State Circuits Conference (ISSCC) , pp. 1522– 1529, 2006

work page 2006

[4] [4]

Resonant clock design for a power-efficient high-volume x86–64 microprocessor,

V . Sathe, S. Arekapudi, C. Ouyang, M. Papaefthymiou, A. Ishii, and S. Naffziger, “Resonant clock design for a power-efficient high-volume x86–64 microprocessor,” in2012 IEEE International Solid-State Circuits Conference, pp. 68–70, 2012

work page 2012

[5] [5]

Design of the two-core x86-64 amd “bulldozer

H. McIntyre, S. Arekapudi, E. Busta, T. Fischer, M. Golden, A. Horiuchi, T. Meneghini, S. Naffziger, and J. Vinh, “Design of the two-core x86-64 amd “bulldozer” module in 32 nm soi cmos,” IEEE Journal of Solid- State Circuits, vol. 47, no. 1, pp. 164–176, 2012

work page 2012

[6] [6]

High- performance clock mesh optimization,

M. Guthaus, X. Hu, G. Wilke, G. Flache, and R. Reis, “High- performance clock mesh optimization,” ACM Transactions on Design Automation of Electronic Systems (TODAES) , 2012

work page 2012

[7] [7]

Meshworks: an efficient framework for planning, synthesis and optimization of clock mesh networks,

A. Rajaram and D. Z. Pan, “Meshworks: an efficient framework for planning, synthesis and optimization of clock mesh networks,” in Asia and South Pacific Design Automation Conference (ASP-DAC) , pp. 250– 257, 2008

work page 2008

[8] [8]

Clock mesh synthesis method using the earth mover’s distance under transformations,

Y . Teng and B. Taskin, “Clock mesh synthesis method using the earth mover’s distance under transformations,” in 2012 IEEE 30th International Conference on Computer Design (ICCD) , pp. 121–126, 2012

work page 2012