Pith. sign in

REVIEW 2 major objections 5 minor 52 references

Fovea: Physical-Implication-Aware Wafer-Scale DSE with Decision-Domain-Guided Cross-Fidelity Refinement

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Fovea claims that a disagreement-bounded Decision Domain, estimated from 10% paired calibration, contains the designated-reference optimum whenever the bound is valid, and that it recovered that optimum in all 70 evaluated workload/design-s

desk verdict A credible, honestly-scoped wafer-scale DSE paper with a correct conditional containment theorem; the headline 100% recovery is empirical and rests on a sampled bound, so it should be reframed as such. read the letter →

arxiv 2608.03285 v1 pith:MBUA5MW3 submitted 2026-08-04 cs.AR

Jinxi Li , Huizheng Wang , Jinyi Deng , Yang Hu , Shouyi Yin This is my paper
classification cs.AR
keywords wafer-scalecomputingdesign-spaceexplorationmulti-fidelityrefinementphysicalimplicationdecisiondomaincross-fidelitydisagreementLLMtrainingreference-optimalrecovery
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 aims to make wafer-scale design-space exploration both affordable and safe, replacing the usual fixed shortlist with a disagreement-driven retention rule. It argues that low-cost analytical evaluation alone can miss the design that a detailed reference evaluator would choose, while exhaustive reference evaluation is prohibitively expensive. Fovea's central claim is that a workload- and space-specific bound on analytical-to-reference disagreement defines a Decision Domain that is guaranteed to contain the reference optimum whenever the bound is valid. Empirically, with only 10% paired calibration, Fovea recovered the exhaustive reference optimum in all 70 tested workload/design-space pairs, at 4.13x average and 7.80x maximum end-to-end speedup.

What carries the argument

The central object is the Decision Domain with retain cutoff $R(\epsilon)=\frac{1-\epsilon}{1+\epsilon}P_L(d^*_L)$. Low-cost scores are converted into reference-consistent intervals $[P_L/(1+\epsilon),\,P_L/(1-\epsilon)]$; a candidate is retained if its optimistic reference performance can still reach the pessimistic bound of the low-cost optimum. The containment proof rests on the chain $P_L(d_R^*)/(1-\epsilon)\ge P_R(d_R^*)\ge P_R(d_L^*)\ge P_L(d_L^*)/(1+\epsilon)$.

What would settle it

Run exhaustive reference evaluation on a held-out or new design space, compute the true full-domain disagreement bound, and compare it with the 10%-sample bound. If any candidate below the retain cutoff has reference performance above the best retained candidate, or if the true maximum disagreement crosses the retention boundary, the containment claim fails for that pair.

Watch

Extended reading notes

Core claim

Under a valid domain-wide disagreement bound, the Decision Domain contains the designated-reference optimum. Concretely, for candidate $d$ with low-cost score $P_L(d)$ and reference score $P_R(d)$, define relative disagreement $e(d)=|P_L(d)/P_R(d)-1|$ and $\epsilon=\max_{d\in D} e(d)$. If $0\le\epsilon<1$, every reference-optimal candidate must satisfy $P_L(d)\ge \frac{1-\epsilon}{1+\epsilon}P_L(d^*_L)$, where $d^*_L$ is the low-cost optimum. Thus, evaluating only the retained Decision Domain and choosing the best reference-scored candidate recovers the exhaustive reference optimum. The implementation estimates $\epsilon$ from 10% paired calibration and, across ten reference-verifiable desig

Load-bearing premise

The load-bearing premise is that the maximum disagreement measured on a 10% calibration sample stands in for the true worst-case disagreement over the whole design space; if a trimmed candidate hides a larger disagreement, Fovea can silently return a design that is not the reference optimum.

Editorial extensions

If this is right

  • With a valid disagreement bound, no reference-optimal candidate is trimmed, so the design returned after selective reference evaluation is exactly the exhaustive reference optimum.
  • The retained fraction is determined by measured disagreement and the low-cost score distribution, not by a preset top-k; tightening the disagreement bound can only shrink the Decision Domain.
  • At the evaluated operating point, Fovea used 20.42% of the reference-evaluation budget on average and still recovered the reference optimum in all 70 workload/design-space pairs.
  • The physical-implication-aware formulation removes 86.4% of area-feasible candidates on average, including 29.4% of analytically top-10% candidates, meaning analytical-only screening would often evaluate physically invalid designs.
  • Workload-dependent reference-optimal designs differ across LLM-training workloads, so workload-specific exploration matters; Fovea's method is reusable rather than a fixed wafer template.

Reading between the lines

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

  • Extension: The same interval-containment recipe should transfer to any low-cost/reference evaluator pair with a valid disagreement bound, e.g., cycle-approximate simulation versus RTL simulation or RTL versus silicon, though the paper only demonstrates the analytical-versus-reference pair it exercises.
  • Extension: The guarantee is conditional on a full-domain bound that is itself estimated from a sample; on a new design space one must revalidate the sampling rate or monitor underestimation, since a boundary-crossing underestimate would silently return a suboptimal design.
  • Extension: Because the paper treats the designated reference backend, not silicon, as ground truth, the 'reference optimum' is relative to that backend; if the reference backend itself differs from silicon, the same containment argument applies one level down.
  • Extension: The paper's baselines receive a fixed 25% runtime budget, slightly above Fovea's measured mean; a per-pair equal-budget comparison could shift baseline recovery rates, though Fovea's exact-recovery result is already at a lower realized cost.
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

2 major / 5 minor

Summary. Fovea proposes a two-stage wafer-scale design-space exploration methodology. Stage I performs physical-implication-aware design-space formulation: it maps die, tiling, compute, memory, interconnect, and D2D choices through reticle, area, boundary-access, and placement constraints, then applies two local reductions that the paper claims are exact. Stage II performs Decision-Domain-guided cross-fidelity refinement: a low-cost analytical evaluator scores the full formed space; paired in-domain calibration estimates evaluator disagreement, which defines a retention cutoff and a Decision Domain C(ε); candidates in C(ε) receive designated-reference evaluation. Section V-B proves that, under a valid full-domain disagreement bound ε, C(ε) contains the designated-reference optimum. The evaluated implementation uses a sampled maximum ε_hat from a 10% paired calibration sample and reports recovery of the exhaustive reference optimum in all 1400 runs across 70 workload/design-space pairs, with 4.13x average and 7.80x maximum end-to-end speedup over exhaustive reference evaluation. The paper also reports a gem5 Garnet cross-validation of the reference backend and compares against simulated annealing, Theseus, and Polaris under a fixed 25% reference budget.

Significance. If the empirical claims are taken at face value, Fovea contributes a reusable, confirmation-centric alternative to search-centric multi-fidelity DSE. The conditional containment theorem in Eq. (11) is correct and parameter-free: the chain P_L(d*_R)/(1-ε) ≥ P_R(d*_R) ≥ P_R(d*_L) ≥ P_L(d*_L)/(1+ε) is valid, and the Decision Domain is not a fitted construct. The evaluation is unusually thorough in using exhaustive reference evaluation for 70 workload/design-space pairs and in cross-validating the reference backend against gem5 Garnet. The physical-implication-aware formulation and its ablation studies are also valuable. The main limitation is that the implementation replaces the theorem's full-domain bound with a sampled maximum, and the reported recovery metric does not directly validate the Decision Domain's containment property; additional analysis is needed to separate genuine domain retention from calibration-sample rescue. If that analysis confirms the current interpretation, the paper's central contribution is significant.

major comments (2)
  1. [Section VI-D / VI-F, Eq. (10), Table V] The containment theorem in Section V-B is conditional on a valid full-domain disagreement bound ε over all d ∈ D, but the implementation uses ε_hat = max_{d∈S} e(d) from a sampled calibration set S, and Section VI-D concedes that ε_hat 'may underestimate the full-domain disagreement bound.' The reported 100% recovery in Table V is not a direct validation of the Decision Domain's containment property, because final selection considers all reference-evaluated candidates in C(ε_hat) ∪ S (Section VI-D). If the designated-reference optimum d*_R is outside C(ε_hat) but happens to be in S, recovery succeeds without C(ε_hat) actually containing d*_R. The retrospective statement that the largest observed underestimation was 1.47 percentage points reports magnitude, not decision impact. Since the paper's central claim is that the Decision Domain contains the designated-reference optimum under a va
  2. [Section IV-E, Fig. 4] The 'same-footprint local strict-dominance reduction' is described as exact and removes a configuration B when A has the same modeled footprint and no less capability in the compared compute, memory, and D2D dimensions. The justification is that the evaluators are monotonic in these capability dimensions. This is a load-bearing assumption for Stage I, but no empirical verification is reported using the existing exhaustive reference corpus. If the monotonicity assumption fails in some corner--for example, through congestion, memory interference, or interconnect effects not captured by the compared dimensions--then a removed configuration could be reference-optimal in the pre-reduction space. Because the exhaustive reference corpus already exists, please verify on the ten reference-verifiable spaces that no removed configuration is reference-optimal, or explicitly qualify the 'exact' claim
minor comments (5)
  1. [Abstract / Section I] There are missing spaces in the abstract and introduction: 'performsPhysical-implication-aware design-space formulationto' and similar. Please proofread for spacing and formatting artifacts.
  2. [Section V-B] The text uses placeholder-like characters 'candidate⌈' and '⌈cannot' in the Decision Domain formation paragraph. These appear to be a LaTeX or conversion artifact and should be fixed.
  3. [Section VI-D] The sampling-based estimate is written as 'ˆϵ' and the text states 'We therefore treatˆϵas a practical disagreement estimate.' Please ensure all inline math renders consistently.
  4. [Section VI-D / Table V] The definition of 'recovery' should be stated explicitly in the main text: recovery currently means the returned design has equal normalized performance to the exhaustive reference optimum, regardless of whether it was found through C(ε_hat) or through S. This is not a flaw per se, but it should be defined in the caption or immediately before Table V given its central role.
  5. [Section VI-B] The gem5 Garnet cross-validation covers only four workload/design-space settings. This is a reasonable start, but the sentence 'This result supports decision-level consistency' should be tempered by noting the limited coverage.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Decision-Domain containment proof is a parameter-free conditional derivation, and the empirical recovery record is measured against exhaustive reference evaluation rather than derived from fitted inputs.

full rationale

The central derivation (Section V-B, Eqs. 8-11) is self-contained and parameter-free: given a valid full-domain disagreement bound ε = max_d e(d), the interval bounds of Eq. 9 imply P_L(d*_R) ≥ ((1-ε)/(1+ε)) P_L(d*_L), so d*_R ∈ C(ε). No fitted constant, calibration sample, or self-citation enters this proof; the contained optimum is not used to define the Decision Domain. The implementation uses a sampled maximum ε_hat = max_{d∈S} e(d) (Section VI-D), and the paper explicitly concedes it 'may underestimate the full-domain disagreement bound.' This is an acknowledged validity/robustness gap between the conditional theorem and the sampled estimator, not a circular reduction: the 100% recovery in Table V is a measured outcome over 1,400 held-out calibration draws against the exhaustive reference corpus, not a consequence of the theorem. The 10% calibration rate is selected by leave-one-design-space-out validation and is the maximum fold-selected rate; because enlarging S monotonically enlarges both the calibration sample and C(ε_hat), this choice is conservative rather than an overfit. The final selection over C(ε_hat) ∪ S (Section VI-F) can in principle recover the optimum through the random sample even if C(ε_hat) excludes it, but this is a possible confound in the empirical attribution, not a definitional equivalence. Self-citations appear only as background/context (e.g., wafer-scale surveys, simulators) and are not load-bearing for the containment argument or the recovery claim.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

Fovea introduces no physical entities, forces, or conserved quantities; the Decision Domain is a defined mathematical set with a containment proof, not a postulated entity. The real cost of the method sits elsewhere: one hyperparameter (sampling rate) fitted to the success metric, one unverifiable domain assumption (bound validity), and a set of hand-set physical parameters (τ_D2D, τ_edge, area budgets) whose sensitivity is unexamined. The formal math is minimal and correct; the empirical burden is carried by the calibration estimate.

free parameters (2)
  • calibration sampling rate ρ = 10% (fold-selected range 8-10%; maximum selected as conservative operating point)
    Selected via leave-one-design-space-out as the smallest rate with no observed recovery failure; the selection objective is the headline success metric, so success at the selected rate is partly a selection outcome.
  • sampled disagreement bound ε_hat = per-pair max of e(d) over the 10% calibration sample; max observed full-domain underestimation 1.47 percentage points
    Estimated from calibration data rather than proven; it parameterizes the retain cutoff and the Decision Domain. The theorem's validity assumption is stated in terms of the unmeasured full-domain version of this quantity.
assumptions (5)
  • domain assumption Evaluators are monotonic in compared capability dimensions within a same-footprint, same-tiling, same-placement class, making the strict-dominance reduction exact.
    Invoked in Section IV-E to justify removing dominated configurations; plausible for fixed DP+TP LLM training but not verified.
  • domain assumption Modeled feasibility checks (reticle, D2D capability envelope, boundary placement) correctly capture real implementation constraints, with hand-set τ_edge parameters (0.20mm corner keep-out, 10% side reservation, 20μm spacing).
    Sections IV-B, IV-D, VI-A; no sensitivity analysis on these parameters is reported, yet they determine which designs are candidates at all.
  • domain assumption The ASTRA-sim analytical and ASTRA-sim+ns-3 reference evaluators faithfully model wafer-scale system performance; ns-3 output is treated as decision ground truth.
    Section II-D states the reference is 'the decision reference used in this study rather than silicon ground truth'; independent cross-validation against gem5 Garnet covers only 4 of 70 pairs (Section VI-B).
  • domain assumption A sampled maximum of e(d) over 10% of candidates approximates the full-domain disagreement bound closely enough to preserve retention decisions.
    Section VI-D admits ε_hat 'may underestimate the full-domain disagreement bound'; the Section V-B theorem requires a valid bound over the complete space, which is never verified.
  • standard math The interval arithmetic and elementary inequalities in Section V (Eqs 8 to 11) are valid for 0 ≤ ε < 1.
    I verified the containment chain P_L(d*_R)/(1-ε) ≥ P_R(d*_R) ≥ P_R(d*_L) ≥ P_L(d*_L)/(1+ε), which yields the retain condition P_L(d*_R) ≥ R(ε).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fovea: Physical-Implication-Aware Wafer-Scale DSE with Decision-Domain-Guided Cross-Fidelity Refinement." pith.science (2026). https://pith.science/paper/MBUA5MW3

@misc{pith2026260803285,
  author       = {Pith},
  title        = {Pith review of: Fovea: Physical-Implication-Aware Wafer-Scale DSE with Decision-Domain-Guided Cross-Fidelity Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBUA5MW3}},
  note         = {Machine review of arXiv:2608.03285}
}
read the original abstract

Modern pre-silicon design-space exploration (DSE) follows a coarse-to-fine workflow: low-cost evaluators screen candidate spaces, while detailed evaluation is reserved for a shortlist. Wafer-scale systems strain both stages. Architecture choices induce coupled physical implications for reticle compliance, wafer tiling, die area, D2D capability, boundary access, and placement, so the design space cannot be treated as an unconstrained Cartesian product. Meanwhile, detailed evaluation is too expensive to cover the resulting space, whereas analytical-to-reference ranking inversions make a fixed shortlist unreliable. We present Fovea, a reusable methodology for workload-specific wafer architecture selection rather than a fixed wafer template. Fovea first performs physical-implication-aware design-space formulation to construct a distinct modeled-feasible space while preserving cross-dimensional trade-offs and applying only evaluator-preserving local reductions. It then performs Decision-Domain-guided cross-fidelity refinement. Paired in-domain calibration estimates workload- and space-specific analytical-to-reference disagreement, which parameterizes reference-consistent performance intervals and induces a Decision Domain for selective designated-reference evaluation. Under a valid domain-wide disagreement bound, this domain contains the designated-reference optimum; the sampling-based implementation is evaluated empirically on exhaustive-reference design spaces. Across ten reference-verifiable design spaces and seven LLM-training workloads, Fovea with 10% paired calibration recovers the exhaustive designated-reference optimum in all 70 evaluated pairs while achieving 4.13x average and 7.80x maximum end-to-end speedup.

Figures

Figures reproduced from arXiv: 2608.03285 by the authors.

Figure 1
Figure 1. Wafer-scale DSE strains two assumptions of conventional multi-stage [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Wafer-scale system model, coupled physical implications, and the tension between evaluation cost and decision reliability (a) Fovea models a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of Fovea. Physical-implication-aware formulation con [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Physical-implication-aware design-space formulation. Fovea enumerates reticle-compliant die outlines and compatible wafer tilings, reserves non [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Decision-Domain-guided cross-fidelity refinement. (a) Low-cost and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Disagreement–Decision-Domain diagnostic. For a fixed low-cost score distribution, (a) tightening the disagreement bound raises the low-cost-score cutoff [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Physical-formulation ablation across the ten reference-verifiable design [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Stage-I construction scale and structure of the modeled-feasible design spaces. The left panel shows the final design-space size as component-library [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: End-to-end Fovea speedup over Exhaustive Reference. Each bar [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Workload-dependent reference-optimal designs. (a) Normalized com [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 48 canonical work pages

  1. [1]

    Garnet: A detailed on-chip network model inside a full-system simulator,

    N. Agarwal, T. Krishna, L.-S. Peh, and N. K. Jha, “Garnet: A detailed on-chip network model inside a full-system simulator,” in2009 IEEE International Symposium on Performance Analysis of Systems and Software, 2009, pp. 33–42

    work page 2009

  2. [2]

    vTrain: A sim- ulation framework for evaluating cost-effective and compute-optimal large language model training,

    J. Bang, Y . Choi, M. Kim, Y . Kim, and M. Rhu, “vTrain: A sim- ulation framework for evaluating cost-effective and compute-optimal large language model training,” in2024 57th IEEE/ACM International Symposium on Microarchitecture (MICRO), 2024, pp. 153–167

    work page 2024

  3. [3]

    Kite: A family of heterogeneous interposer topologies enabled via accurate interconnect modeling,

    S. Bharadwaj, J. Yin, B. Beckmann, and T. Krishna, “Kite: A family of heterogeneous interposer topologies enabled via accurate interconnect modeling,” inProceedings of the 57th ACM/IEEE Design Automation Conference, 2020, pp. 1–6

    work page 2020

  4. [4]

    The gem5 Simulator,

    N. Binkert, B. Beckmann, G. Black, S. K. Reinhardt, A. Saidi, A. Basu, J. Hestness, D. R. Hower, T. Krishna, S. Sardashti, R. Sen, K. Sewell, M. Shoaib, N. Vaish, M. D. Hill, and D. A. Wood, “The gem5 Simulator,”ACM SIGARCH Computer Architecture News, vol. 39, no. 2, pp. 1–7, 2011

    work page 2011

  5. [5]

    T.-Y . J. Chang, Y .-H. Chen, W.-M. Chan, H. Cheng, P.-S. Wang, Y . Lin, H. Fujiwara, R. Lee, H.-J. Liao, P.-W. Wang, G. Yeap, and Q. Li, “A 5-nm 135-mb SRAM in EUV and high-mobility-channel FinFET technology with metal coupling and charge-sharing write-assist circuitry schemes for high-density and low-v min applications,”IEEE Journal of Solid-State Circu...

    work page 2021

  6. [6]

    CHASE: A chiplet architecture simulation and exploration framework with decoupled multi-fidelity optimization,

    S. Chen, H. Zhang, J. Zhai, and B. Yu, “CHASE: A chiplet architecture simulation and exploration framework with decoupled multi-fidelity optimization,” inProceedings of the 2026 International Symposium on Physical Design. ACM, 2026, pp. 153–161

    work page 2026

  7. [7]

    Eval- uating the impacts of reticle requirements in semiconductor wafer fab- rication,

    S. de Diaz, J. Fowler, M. Pfund, G. Mackulak, and M. Hickie, “Eval- uating the impacts of reticle requirements in semiconductor wafer fab- rication,”IEEE Transactions on Semiconductor Manufacturing, vol. 18, no. 4, pp. 622–632, 2005

    work page 2005

  8. [8]

    Proteus: Simulating the performance of distributed DNN training,

    J. Duan, X. Li, P. Xu, X. Zhang, S. Yan, Y . Liang, and D. Lin, “Proteus: Simulating the performance of distributed DNN training,” IEEE Transactions on Parallel and Distributed Systems, vol. 35, no. 10, pp. 1867–1878, 2024

    work page 2024

Show all 52 references
  1. [9]

    Explainable fuzzy neural network with multi-fidelity reinforcement learning for micro- architecture design space exploration,

    H. Fan, Y . Wang, S. Li, T. Liang, and W. Zhang, “Explainable fuzzy neural network with multi-fidelity reinforcement learning for micro- architecture design space exploration,” inProceedings of the 61st ACM/IEEE Design Automation Conference. ACM, 2024, pp. 7:1–7:6

    2024

  2. [11]

    System–technology co-optimization for dense edge archi- tectures using 3-D integration and nonvolatile memory,

    L. M. Giacomini Rocha, M. Naeim, G. Paim, M. Brunion, P. Venugopal, D. Milojevic, J. Myers, M. Badaroglu, M. Verhelst, J. Ryckaert, and D. Biswas, “System–technology co-optimization for dense edge archi- tectures using 3-D integration and nonvolatile memory,”IEEE Journal on Ex...

    2024

  3. [12]

    Monad: Towards cost-effective specialization for chiplet-based spatial accelerators,

    X. Hao, Z. Ding, J. Yin, Y . Wang, and Y . Liang, “Monad: Towards cost-effective specialization for chiplet-based spatial accelerators,” in 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), 2023, pp. 1–9

    2023

  4. [13]

    Wafer- scale computing: Advancements, challenges, and future perspectives,

    Y . Hu, X. Lin, H. Wang, Z. He, X. Yu, J. Zhang, Q. Yang, Z. Xu, S. Guan, J. Fang, H. Shang, X. Tang, X. Dai, S. Wei, and S. Yin, “Wafer- scale computing: Advancements, challenges, and future perspectives,” IEEE Circuits and Systems Magazine, vol. 24, no. 1, pp. 52–81, 2024

    2024

  5. [14]

    ISPD 2020 physical mapping of neural networks on a wafer-scale deep learning accelerator,

    M. James, M. Tom, P. Groeneveld, and V . Kibardin, “ISPD 2020 physical mapping of neural networks on a wafer-scale deep learning accelerator,” inProceedings of the 2020 International Symposium on Physical Design. ACM, 2020, pp. 145–149

    2020

  6. [15]

    Union: A unified HW–SW co- design ecosystem in MLIR for evaluating tensor operations on spatial accelerators,

    G. Jeong, G. Kestor, P. Chatarasi, A. Parashar, P.-A. Tsai, S. Rajaman- ickam, R. Gioiosa, and T. Krishna, “Union: A unified HW–SW co- design ecosystem in MLIR for evaluating tensor operations on spatial accelerators,” in2021 International Conference on Parallel Architectures ...

    2021

  7. [16]

    FireSim: FPGA- accelerated cycle-exact scale-out system simulation in the public cloud,

    S. Karandikar, H. Mao, D. Kim, D. Biancolin, A. Amid, D. Lee, N. Pemberton, E. Amaro, C. Schmidt, A. Chopra, Q. Huang, K. Kovacs, B. Nikolic, R. H. Katz, J. Bachrach, and K. Asanovic, “FireSim: FPGA- accelerated cycle-exact scale-out system simulation in the public cloud,” in2...

    2018

  8. [17]

    Impact of RoCE congestion control policies on distributed training of DNNs,

    T. Khan, S. Rashidi, S. Sridharan, P. Shurpali, A. Akella, and T. Krishna, “Impact of RoCE congestion control policies on distributed training of DNNs,” in2022 IEEE Symposium on High-Performance Interconnects (HOTI), 2022, pp. 39–48

    2022

  9. [18]

    Optimization by simulated annealing,

    S. Kirkpatrick, C. D. Gelatt, Jr., and M. P. Vecchi, “Optimization by simulated annealing,”Science, vol. 220, no. 4598, pp. 671–680, 1983

    1983

  10. [19]

    MAESTRO: A data-centric approach to understand reuse, performance, and hardware cost of DNN mappings,

    H. Kwon, P. Chatarasi, V . Sarkar, T. Krishna, M. Pellauer, and A. Parashar, “MAESTRO: A data-centric approach to understand reuse, performance, and hardware cost of DNN mappings,”IEEE Micro, vol. 40, no. 3, pp. 20–29, 2020

    2020

  11. [20]

    The path to successful wafer-scale integration: The Cerebras story,

    G. Lauterbach, “The path to successful wafer-scale integration: The Cerebras story,”IEEE Micro, vol. 41, no. 6, pp. 52–57, 2021

    2021

  12. [21]

    ReThermal: Co-design of thermal-aware static and dynamic scheduling for LLM training on liquid-cooled wafer-scale chips,

    C. Li, H. Wang, J. Liu, J. Liu, Z. Yue, X. Li, S. Jiang, J. Deng, Y . Hu, and S. Yin, “ReThermal: Co-design of thermal-aware static and dynamic scheduling for LLM training on liquid-cooled wafer-scale chips,” in 2026 IEEE International Symposium on High Performance Computer Ar...

    2026

  13. [22]

    SPAC: Automating FPGA-based network switches with protocol adaptive customization,

    G. Li, Y . Cao, L. H. L. Ng, A. Charlton, Q. Wang, W. Punter, P. Papaphilippou, C. Guo, H. Fan, W. Luk, S. P. Amarasinghe, and A. Brahmakshatriya, “SPAC: Automating FPGA-based network switches with protocol adaptive customization,” in2026 IEEE 34th Annual International Symposi...

    2026

  14. [23]

    Swift or exact? boosting efficient microarchitecture DSE via multi-fidelity partial-order prediction,

    H. Liu, H. Geng, Z. He, Q. Sun, and C. Zhuo, “Swift or exact? boosting efficient microarchitecture DSE via multi-fidelity partial-order prediction,” in2025 62nd ACM/IEEE Design Automation Conference (DAC), 2025, pp. 1–7

    2025

  15. [24]

    Multi-fidelity optimization for high-level syn- thesis directives,

    C. Lo and P. Chow, “Multi-fidelity optimization for high-level syn- thesis directives,” in2018 28th International Conference on Field Programmable Logic and Applications (FPL), 2018, pp. 272–279

    2018

  16. [25]

    A 0.297-pj/bit 50.4-gb/s/wire inverter-based short-reach simultaneous bi-directional transceiver for die-to-die interface in 5-nm cmos,

    Y . Nishi, J. W. Poulton, W. J. Turner, X. Chen, S. Song, B. Zimmer, S. G. Tell, N. Nedovic, J. M. Wilson, W. J. Dally, and C. T. Gray, “A 0.297-pj/bit 50.4-gb/s/wire inverter-based short-reach simultaneous bi-directional transceiver for die-to-die interface in 5-nm cmos,”IEEE...

    2023

  17. [26]

    Bunch of wires (bow) phy specification,

    Open Compute Project ODSA BoW Workstream, “Bunch of wires (bow) phy specification,” 2023, draft Version 1.9d

    2023

  18. [27]

    OpenAI Partners with Cerebras,

    OpenAI, “OpenAI Partners with Cerebras,” https://openai.com/index/ cerebras-partnership/, Jan. 2026, accessed: Aug. 1, 2026

    2026

  19. [28]

    MuchiSim: A simulation framework for design exploration of multi-chip manycore systems,

    M. Orenes-Vera, E. Tureci, M. Martonosi, and D. Wentzlaff, “MuchiSim: A simulation framework for design exploration of multi-chip manycore systems,” in2024 IEEE International Symposium on Performance Anal- ysis of Systems and Software (ISPASS), 2024, pp. 48–60

    2024

  20. [29]

    Architecting wafer-scale processors—a GPU case study,

    S. Pal, D. Petrisko, M. Tomei, S. S. Iyer, P. Gupta, and R. Kumar, “Architecting wafer-scale processors—a GPU case study,” in2019 IEEE International Symposium on High Performance Computer Architecture (HPCA), 2019, pp. 250–263

    2019

  21. [30]

    Timeloop: A systematic approach to DNN accelerator evaluation,

    A. Parashar, P. Raina, Y . S. Shao, Y .-H. Chen, V . A. Ying, A. Mukkara, R. Venkatesan, B. Khailany, S. W. Keckler, and J. S. Emer, “Timeloop: A systematic approach to DNN accelerator evaluation,” in2019 IEEE International Symposium on Performance Analysis of Systems and Soft...

    2019

  22. [31]

    Application of full-system simulation in exploratory system design and development,

    J. L. Peterson, P. J. Bohrer, L. Chen, E. N. Elnozahy, A. Gheith, R. H. Jewell, M. D. Kistler, T. R. Maeurer, S. A. Malone, D. B. Murrell, N. Needel, K. Rajamani, M. A. Rinaldi, R. O. Simpson, K. Sudeep, and L. Zhang, “Application of full-system simulation in exploratory syste...

    2006

  23. [32]

    Power delivery for high-performance microprocessors—challenges, solutions, and future trends,

    K. Radhakrishnan, M. Swaminathan, and B. K. Bhattacharyya, “Power delivery for high-performance microprocessors—challenges, solutions, and future trends,”IEEE Transactions on Components, Packaging and Manufacturing Technology, vol. 11, no. 4, pp. 655–671, 2021

    2021

  24. [33]

    ASTRA-sim: Enabling SW/HW co-design exploration for distributed DL training platforms,

    S. Rashidi, S. Sridharan, S. Srinivasan, and T. Krishna, “ASTRA-sim: Enabling SW/HW co-design exploration for distributed DL training platforms,” in2020 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), 2020, pp. 81–92

    2020

  25. [34]

    FRED: A wafer-scale fabric for 3d parallel DNN training,

    S. Rashidi, W. Won, S. Srinivasan, P. Gupta, and T. Krishna, “FRED: A wafer-scale fabric for 3d parallel DNN training,” inProceedings of the 52nd Annual International Symposium on Computer Architecture. ACM, 2025, pp. 34–48

    2025

  26. [35]

    Fast stencil-code computation on a wafer-scale processor,

    K. Rocki, D. Van Essendelft, I. Sharapov, R. Schreiber, M. Morrison, V . Kibardin, A. Portnoy, J.-F. Dietiker, M. Syamlal, and M. James, “Fast stencil-code computation on a wafer-scale processor,” inSC20: International Conference for High Performance Computing, Networking, Sto...

    2020

  27. [36]

    Polaris: Multi-fidelity design space exploration of deep learning accelerators,

    C. Sakhuja, C. Hong, and C. Lin, “Polaris: Multi-fidelity design space exploration of deep learning accelerators,”arXiv preprint arXiv:2412.15548, 2024. [Online]. Available: https://arxiv.org/abs/2412. 15548

    2024 arXiv

  28. [37]

    SoW-X: A novel system-on-wafer technology for next generation AI server application,

    P.-C. Shih, A.-J. Su, K.-H. Tam, T.-C. Huang, K. Chuang, and J. Yeh, “SoW-X: A novel system-on-wafer technology for next generation AI server application,” in2025 IEEE 75th Electronic Components and Technology Conference (ECTC), 2025, pp. 1–6

    2025

  29. [38]

    MLCommons Chakra: Advancing performance benchmarking and co-design using standardized execution traces,

    S. Sridharan, A. Balogh, B. M. Beckmann, B. Coutinho, L. Feng, S. Fu, S. Gao, M. Garakani, T. Heo, D. Kanter, J. Ladd, Z. Li, W. Liu, C. Man, D. Mihailescu, S. More, J. Park, A. Ramachandran, V . Ramakrishnaiah, S. Rashidi, V . J. Reddi, P. Sharma, P. Tian, W. Won, H. Wu, H. X...

    2026 arXiv

  30. [39]

    Correlated multi-objective multi-fidelity optimization for HLS directives design,

    Q. Sun, T. Chen, S. Liu, J. Miao, J. Chen, H. Yu, and B. Yu, “Correlated multi-objective multi-fidelity optimization for HLS directives design,” in2021 Design, Automation & Test in Europe Conference & Exhibition (DATE), 2021, pp. 46–51

    2021

  31. [40]

    The microarchitecture of Dojo, Tesla’s exa-scale computer,

    E. Talpes, D. Das Sarma, D. Williams, S. Arora, T. Kunjan, B. Floering, A. Jalote, C. Hsiong, C. Poorna, V . Samant, J. Sicilia, A. K. Nivarti, R. Ramachandran, T. Fischer, B. Herzberg, B. McGee, G. Venkatara- manan, and P. Banon, “The microarchitecture of Dojo, Tesla’s exa-sc...

    2023

  32. [41]

    W ATOS: Efficient LLM training strategies and architecture co- exploration for wafer-scale chip,

    H. Wang, Z. Wang, H. Wang, J. Hou, T. Wei, C. Li, Y . Hu, and S. Yin, “W ATOS: Efficient LLM training strategies and architecture co- exploration for wafer-scale chip,” in2026 IEEE International Sympo- sium on High Performance Computer Architecture (HPCA), 2026, pp. 1–19

    2026

  33. [42]

    Designing spatial architectures for sparse attention: STAR accelerator via cross-stage tiling,

    H. Wang, T. Wei, H. Wang, Z. Wang, X. Tang, Z. Yue, S. Wei, Y . Hu, and S. Yin, “Designing spatial architectures for sparse attention: STAR accelerator via cross-stage tiling,”IEEE Transactions on Computers, vol. 75, no. 3, pp. 1125–1140, 2026

    2026

  34. [43]

    TEMP: A memory efficient physical-aware tensor partition-mapping framework on wafer-scale chips,

    H. Wang, T. Wei, Z. Wang, D. Jiang, Q. Yang, J. Liu, J. Hou, C. Li, J. Deng, Y . Hu, and S. Yin, “TEMP: A memory efficient physical-aware tensor partition-mapping framework on wafer-scale chips,” in2026 IEEE International Symposium on High Performance Computer Architecture (HP...

    2026

  35. [44]

    TMAC: Training- targeted mapping and architecture co-exploration for wafer-scale chips,

    H. Wang, Q. Yang, T. Wei, X. Yu, C. Li, J. Fang, G. Lu, X. Dai, L. Liu, S. Jiang, Y . Hu, S. Yin, and S. Wei, “TMAC: Training- targeted mapping and architecture co-exploration for wafer-scale chips,” Integrated Circuits and Systems, vol. 1, no. 4, pp. 178–195, 2024

    2024

  36. [45]

    MOCAP: Wafer-scale-chip-oriented memory-orchestrated chunked pipelining framework for prefill-only LLM inference,

    Z. Wang, H. Wang, Y . Xiao, H. Zuo, T. Wei, J. Deng, C. Li, Y . Hu, and S. Yin, “MOCAP: Wafer-scale-chip-oriented memory-orchestrated chunked pipelining framework for prefill-only LLM inference,”arXiv preprint arXiv:2606.22968, 2026, accepted to APPT 2026. [Online]. Available:...

    2026 arXiv

  37. [46]

    Spatial-aware orches- tration of LLM attention on waferscale chips,

    T. Wei, H. Wang, Z. Wang, S. Yin, and Y . Hu, “Spatial-aware orches- tration of LLM attention on waferscale chips,” inAdvanced Parallel Processing Technologies: 16th International Symposium, APPT 2025, Athens, Greece, July 13–16, 2025, Proceedings, ser. Lecture Notes in Comput...

    2025

  38. [47]

    ASTRA-sim2.0: Modeling hierarchical networks and disaggregated systems for large-model training at scale,

    W. Won, T. Heo, S. Rashidi, S. Sridharan, S. Srinivasan, and T. Krishna, “ASTRA-sim2.0: Modeling hierarchical networks and disaggregated systems for large-model training at scale,” in2023 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), 20...

    2023

  39. [48]

    WSC-LLM: Efficient LLM service and architecture co- exploration for wafer-scale chips,

    Z. Xu, D. Kong, J. Liu, J. Li, J. Hou, X. Dai, C. Li, S. Wei, Y . Hu, and S. Yin, “WSC-LLM: Efficient LLM service and architecture co- exploration for wafer-scale chips,” inProceedings of the 52nd Annual International Symposium on Computer Architecture. ACM, 2025, pp. 1–17

    2025

  40. [49]

    Apollo: Transferable architecture exploration,

    A. Yazdanbakhsh, C. Angermueller, B. Akin, Y . Zhou, A. Jones, M. Hashemi, K. Swersky, S. Chatterjee, R. Narayanaswami, and J. Laudon, “Apollo: Transferable architecture exploration,” inML for Systems Workshop at NeurIPS, 2020. [Online]. Available: https: //research.google/pub...

    2020

  41. [50]

    G. Yeapet al., “5 nm CMOS production technology platform featur- ing full-fledged EUV and high-mobility-channel FinFETs with densest 0.021µm 2 SRAM cells for mobile SoC and high-performance comput- ing applications,” in2019 IEEE International Electron Devices Meeting (IEDM), 2...

    2019

  42. [51]

    What is DTCO?: An introduction to design–technology co-optimization,

    L. Yuan, “What is DTCO?: An introduction to design–technology co-optimization,” TSMC Technology Blog, Taiwan Semiconductor Manufacturing Company, Jun. 2022, accessed: July 31, 2026. [Online]. Available: https://www.tsmc.com/english/news-events/blog- article-20220615

    2022

  43. [52]

    Theseus: Exploring efficient wafer-scale chip design for large language models,

    J. Zhu, C. Xue, Y . Chen, Z. Wang, C. Zhang, Y . Shen, Y . Chen, Z. Cheng, Y . Jiang, T. Wang, Y . Lin, W. Hu, B. Cui, R. Wang, Y . Liang, and G. Sun, “Theseus: Exploring efficient wafer-scale chip design for large language models,”IEEE Transactions on Computer-Aided Design of...

    2025

  44. [2024]

    Available: https://arxiv.org/abs/2406.03868

    [Online]. Available: https://arxiv.org/abs/2406.03868

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.