Pith. sign in

REVIEW 4 major objections 6 minor 36 references

C2C-Explorer: An Exploration Framework for Chip-to-Chip Interconnect Architectures in LLM Cloud Computing Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read C2C-Explorer claims that a cycle/event simulator plus Bayesian search turns LLM workload traces into chip-to-chip hardware configurations with 2.46–8.23% timing error and 44.1% goodput gains.

desk verdict Useful open-source C2C simulator with a real 4-chip FPGA validation, but the 32-XPU headline numbers depend on scale-transfer that is asserted, not shown. read the letter →

arxiv 2608.08611 v1 pith:VM4LEQP4 submitted 2026-08-09 cs.DC cs.ARcs.CE

Jiayi Li , Di Wu , Qingxu Li , Hongxiao Zhao , Jiaqi Yang , Anjunyi Fan , Wenbin Zhang , Boqiang Wu
show 5 more authors
Shuting Liu Shifeng Fang Jianbo Dong Dimin Niu Bonan Yan
This is my paper · ORCID
classification cs.DCcs.ARcs.CE
keywords chip-to-chipinterconnectLLMworkloadtrafficgenerationhybridcycle-eventsimulationBayesiandesignspaceexplorationscale-upflowcompletiontimeAXImodelingcredit-basedcontrol
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 claims that the design of chip-to-chip (C2C) interconnect hardware for large-language-model systems can be automated end to end from workload traces. It builds a pipeline that turns LLM collective-communication traces into AXI-level traffic, simulates that traffic through a hybrid cycle-accurate/event-driven interconnect model, scores the result, and searches the hardware configuration space with Bayesian optimization. The authors argue that, if the approach holds at scale, it closes the gap between LLM software workloads and C2C hardware parameters such as frame size, virtual-channel count, and scheduling policy. On a four-chip FPGA prototype the simulator matches measured flow-completion timing to within 2.46–8.23%, and on a 32-chip inference workload the search reports 44.1% higher goodput and 98.4% lower buffer usage than the worst feasible designs it considered.

What carries the argument

The load-bearing mechanism is the hybrid timing model: cycle-accurate C2C port stages, coupled through producer-consumer FIFOs, for the AXI and MAC datapath, paired with an event-driven switch stage that is an input-queued crossbar with per-source-per-destination virtual output queues, round-robin arbitration, and approximated cut-through. This split keeps per-frame serialization, credit backpressure, and virtual-channel occupancy cycle-exact where they matter, while replacing cycle-by-cycle switch simulation with event scheduling to gain the reported speedup. The traffic generator's dual flow-control semantics—application-level bandwidth-delay-product sliding window plus link-level credit-based backpressure—is the object that connects LLM traces to the simulator.

What would settle it

Run the same 32-chip all-to-all inference trace on a real 32-chip prototype with the reported link parameters, and compare measured per-flow completion times and total goodput against the simulator's predictions; if the end-to-end error grows well beyond the 2.46–8.23% band or the 44.1% goodput gain does not materialize, the scale extrapolation fails.

Watch

Extended reading notes

Core claim

C2C-Explorer's central claim is that a workload-to-hardware co-design loop can be made accurate, fast, and searchable enough to be useful for real LLM systems. The loop starts with a traffic generator that maps LLM point-to-point traces onto per-port C2C flows and emits AXI write bursts under a sliding-window plus credit-based flow control; then a simulator models each C2C port cycle-accurately through a seven-stage pipeline while treating the switch stage as an event-driven input-queued crossbar with virtual output queues. The authors report that this hybrid model reproduces end-to-end timing on an FPGA prototype within 2.46–8.23% across one-to-all, all-to-one, and all-to-all patterns, and that it runs up to 7.8× faster than a pure cycle-accurate model at up to 512 chips. Finally, an adaptive Bayesian search prunes the design space using hardware-feasibility constraints and converges within about 20 evaluations; applied to a 32-chip mixture-of-experts inference workload, the best configurations improve goodput by 44.1% and cut buffer provisioning by 98.4% relative to the worst feasible designs.

Load-bearing premise

The accuracy validation is done on a single four-chip, one-switch prototype, and the paper's headline design-space results assume that the same timing fidelity carries over to 32-chip and larger systems that were never checked against hardware.

Editorial extensions

If this is right

  • Designers can prune the C2C configuration space before simulation: the chunk_size ≥ 2 × MAC_frame constraint alone cuts the feasible set from 2394 to 1152 configurations.
  • The search converges within about 20 simulator evaluations, so a designer with a single workload trace can explore hundreds of configurations cheaply instead of relying on hand-tuned rules.
  • Parameter sensitivity is workload-dependent and non-monotonic, so the right C2C configuration for one model or parallelization strategy is not necessarily right for another.
  • The hybrid simulator's speedup grows with parallelism, up to 7.8× at 512 chips, making hardware-level exploration at scale-out sizes more practical.
  • Metric-guided scoring lets the same framework optimize for different objectives, such as throughput-heavy inference versus tail-latency-sensitive serving, by changing the weights.

Reading between the lines

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

  • Because the accuracy check stops at four chips, the most direct next test is a mid-scale 16-to-64 chip hardware comparison; until then the 32-chip design-space results should be read as conditioned on the hybrid switch model's fidelity under contention.
  • The framework's components are largely standard—AXI bursts, credit flow control, virtual-output-queue crossbar arbitration—so the same loop may transfer to other scale-up interconnect standards if port parameters and PHY delays are substituted; that generality is plausible but unverified.
  • The reported workload dependence suggests a future where interconnect configuration is tuned per model or per deployment, and the framework could be used to produce configuration runbooks automatically.
  • The 98.4% buffer reduction implies that most of the default buffer provisioning in current designs is safety margin; if confirmed at scale, this could translate directly into silicon area and power savings.
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

4 major / 6 minor

Summary. The paper proposes C2C-Explorer, an open-source framework for exploring chip-to-chip (C2C) interconnect architectures in large-scale LLM computing systems. It combines an LLM-workload-driven traffic generator (built on SimAI traces), a hybrid cycle-accurate/event-driven C2C simulator supporting switch and full-mesh topologies, a metric-guided evaluator, and an adaptive Bayesian design-space-exploration engine. The central claims are: (i) the simulator is validated against a 4-XPU, 1-switch 400 Gbps FPGA prototype and achieves 2.46--8.23% mean end-to-end timing error across One-to-All, All-to-One, and All-to-All patterns; (ii) the hybrid model accelerates simulation by up to 7.8x over a pure cycle-accurate baseline; and (iii) on a 32-XPU DeepSeek-R1-671B inference workload, the framework identifies configurations that improve goodput by 44.1% and reduce buffer usage by 98.4% relative to the worst feasible pruned designs. The paper also reports parameter-sensitivity studies and convergence of the Bayesian search within about 20 iterations.

Significance. If the claims hold, C2C-Explorer would be a genuinely useful open-source tool: it addresses a real gap between LLM-level workload simulation and hardware-level C2C interconnect simulation, and it ships an artifact (the GitHub repository) that can be independently exercised. The small-scale hardware validation is a concrete strength, as is the explicit modeling of AXI transactions, credit-based flow control, and MAC-layer packetization. The hybrid cycle/event modeling idea is pragmatic and the speedup measurements are plausible. However, the significance of the headline results is currently limited by two issues: the accuracy validation is only at four chips while the main design-space results are at 32 chips, and the reported DSE improvements are relative to worst feasible designs rather than to a meaningful baseline. These are fixable with additional experiments and clearer framing, so I do not see them as fatal; they do require substantive revision.

major comments (4)
  1. [Section 4.1 and Section 6] The accuracy validation is performed on a single 4-XPU, 1-switch FPGA prototype, but the headline DSE results (Table 3) are obtained for a 32-XPU system, and the scalability experiments in Section 4.2 reach 512 XPUs. The event-driven switch stage S4 is an input-queued crossbar with per-(src,dst,vc) VOQs, round-robin arbitration, and approximated cut-through; its timing error could plausibly grow with port count and with all-to-all contention. The paper provides no accuracy check of the hybrid model at any scale beyond the 4-XPU prototype, and Figure 5(e-f) compares the hybrid model with the pure cycle-accurate model only in runtime, not in predicted flow completion times. Please add at least one accuracy comparison between the hybrid and pure cycle-accurate models at larger port counts, ideally against a cycle-accurate switch reference or, if available, a larger FPGA prototype, and state how contention and port count affect the error.
  2. [Section 4.1] The text is ambiguous about what the reported 2.46--8.23% errors are measured against. It says the simulator is configured with ETH-X parameters from the ODCC test report [33], that the authors built a 400 Gbps FPGA prototype that 'exposes the same datapath behavior as ETH-X' [32], and then that the mean error is computed 'relative to the ETH-X measurements.' Please clarify whether the reference curve is (a) measurements taken on the authors' own FPGA prototype, or (b) numbers transcribed from the ODCC/ETH-X public test report. If the latter, the abstract's phrase 'validated against FPGA-based C2C prototypes' is not supported by data presented in this paper, and the validation is external to the authors' prototype. If the former, please state explicitly that the reported errors are for the authors' prototype and explain how the equivalence to ETH-X was established. This distinction is load-bearing for the central accuracy claim.
  3. [Table 3 and abstract] The improvements in Table 3 -- including the '44.1% goodput' and '98.4% buffer reduction' figures quoted in the abstract -- are computed relative to the 'worst feasible designs' after pruning with the hard constraint chunk_size >= 2*MAC_frame. Comparing against the worst feasible design is a weak baseline; the abstract does not state this qualifier, and a reader would naturally interpret the numbers as improvements over a default or state-of-the-art configuration. Please report the improvement relative to a concrete, named baseline (e.g., ETH-X-like defaults, a round-robin/1-VC configuration, or the best configuration found without pruning), and adjust the abstract and conclusion accordingly.
  4. [Section 4.2] The speedup results in Figure 5(e-f) show that the hybrid model sometimes runs slower than the pure cycle-accurate baseline (speedups as low as 0.72x for serialized flows). This is not itself an error, but the paper does not discuss the conditions under which the hybrid model fails to accelerate simulation, nor does it give guidance for when a user should choose the pure cycle-accurate model. Given that the hybrid model also changes the timing semantics of the switch, a brief analysis of where the crossover lies (port count, flow concurrency, message size) would strengthen the practical contribution.
minor comments (6)
  1. [Section 3.1, equation for W] The sliding-window equation is typeset with two expressions joined by a comma: W = ceil(BW*RTT/(C*N*alpha)) and W = ceil((RTT/C)*(BW+Delta)*N*alpha). This is confusing: are these alternative formulas, or are the symbols being redefined? Please use a single definition and define every symbol (especially N*alpha) clearly.
  2. [Section 4.1] Only mean errors over all swept points are reported. Since tail behavior is central to interconnect design, please also report per-pattern worst-case errors and, if possible, a plot or table of per-message-size errors. The 8.23% mean error for All-to-All could hide larger per-point deviations.
  3. [Table 3] The Fairness column contains a '-' entry for the 'combine' row. Please either report the value or state explicitly why it is unavailable.
  4. [Related Work, references [18]-[20]] In Section 2, Garnet is cited as reference [19], but reference [19] is actually the Kite paper. Please correct the citation or add a separate reference for Garnet.
  5. [Section 4.2] Minor wording: '2x Intel Xeon Platinum8358 (1TB main memory)' should read 'on a dual-socket Intel Xeon Platinum 8358 server with 1TB main memory.'
  6. [Abstract and Table 3] The abstract says 'reduce memory by 98.4%' while Table 3 reports 'Buffer Usage.' Since the paper defines buffer usage as NV*CB (a relative provisioning metric), please use consistent terminology and clarify in the text that this is not a physical memory measurement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the simulator's accuracy claim is anchored to an external FPGA prototype and public ODCC/ETH-X reports, and the DSE results are optimization outputs rather than predictions fitted to their own inputs.

full rationale

The paper's load-bearing accuracy claim is externally anchored. In Sec. 4.1, the simulator is configured using public ETH-X technical-report parameters (400 Gbps links, 467 ns switch forwarding delay, 245 ns port processing latency) and then compared against a real 400 Gbps FPGA prototype built by the authors. The reported 2.46%, 4.39%, and 8.23% errors are measured against that hardware baseline, not against the authors' own prior results or against parameters fitted from the simulator. The DSE user study (Sec. 6, Table 3) reports goodput and buffer improvements produced by the Bayesian search over pruned feasible configurations; these are optimization outcomes relative to the worst feasible designs, not predictions that reduce by construction to the evaluator or traffic generator. The only notable self-overlap is SimAI, cited as the source of LLM P2P traces; Qingxu Li appears on both papers. However, SimAI is a separately peer-reviewed, published system (NSDI 2025), and the central claims of C2C-Explorer do not reduce to SimAI's trace fidelity. The unvalidated scale transfer from a 4-XPU prototype to 32- and 512-XPU studies, and the unproven claim that the prototype 'exposes the same datapath behavior as ETH-X', are evidence and correctness limitations, not circular definitions or fitted-input predictions. No quoted equation or parameter is equivalent to a claimed output by construction, and no load-bearing argument rests on a self-citation chain. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper's contribution is an integration of existing components (SimAI traces, SimPy simulation, GP-based Bayesian optimization) with two new pieces of machinery: the P2P-to-AXI traffic mapping with dual flow control, and the hybrid cycle/event timing model. Its external anchors are the ODCC/ETH-X public parameters and the authors' self-built 4-host FPGA prototype; its internal knobs are the evaluation weights, the unquantified window safety factor alpha, and the self-derived pruning rule. No invented physical or protocol entities are introduced; the buffer metric NV*CB is disclosed as a provisioning proxy rather than literal hardware size.

free parameters (3)
  • Evaluation metric weights w_m = goodput 0.40, P50 0.15, P99 0.25, fairness 0.05, buffer 0.15
    Chosen by hand in Sec 3.3 to reflect 'requirements of large-scale inference workloads'; 'optimal' configurations are optimal for this author-defined weighted score, not for an external objective.
  • Sliding-window safety factor alpha = unspecified
    Appears in the window-size formula W = ceil(BW*RTT/(C*N*alpha)) in Sec 3.1; no value is reported, leaving the traffic generator's pacing with an unquantified knob.
  • Pruning rule chunk_size >= 2 * MAC_frame = 2x
    Derived from the goodput saturation plateau measured by the authors in Sec 5.1 (Fig 6a), then applied in Sec 6 as a hard constraint, shrinking the space from 2394 to 1152; an empirical rule fit to the authors' own simulation output.
assumptions (5)
  • domain assumption SimAI-generated P2P traces faithfully represent real LLM communication patterns for DeepSeek-R1, LLaMA3.1, and Qwen3 workloads
    The traffic generator consumes SimAI traces (Sec 3, Sec 6); trace fidelity is taken on trust from the prior tool, whose authors overlap with this paper's team (co-author Qingxu Li).
  • domain assumption ETH-X protocol parameters (400Gbps, 467ns switch forwarding, 245ns port processing) from the ODCC white papers, and the authors' 4-host FPGA prototype, replicate real C2C datapath behavior
    Sec 4.1 configures the simulator from ODCC white papers and assumes the self-built prototype 'exposes the same datapath behavior as ETH-X'; this is the external anchor for the accuracy claim.
  • ad hoc to paper Event-driven switch scheduling preserves cycle-level timing accuracy at scales beyond the validated 4-chip setup
    Sec 3.2 S4 models the switch as an input-queued crossbar with VOQs and round-robin arbitration, with cut-through approximated by overlap; this simplification is the source of the 7.8x speedup but is never validated against hardware at 32-512 chip scale.
  • domain assumption The normalized weighted score S is a valid scalar objective for comparing C2C configurations
    Sec 3.3 constructs S from normalized metrics with weights prioritizing throughput and P99; optimality is relative to this author-defined objective, not to measured end-to-end LLM performance.
  • standard math Standard Bayesian optimization machinery: Gaussian-process surrogate, expected improvement acquisition, Latin hypercube initialization
    Algorithm 1 relies on standard BO results; no new statistical theory is claimed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of C2C-Explorer: An Exploration Framework for Chip-to-Chip Interconnect Architectures in LLM Cloud Computing Systems." pith.science (2026). https://pith.science/paper/VM4LEQP4

@misc{pith2026260808611,
  author       = {Pith},
  title        = {Pith review of: C2C-Explorer: An Exploration Framework for Chip-to-Chip Interconnect Architectures in LLM Cloud Computing Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VM4LEQP4}},
  note         = {Machine review of arXiv:2608.08611}
}
abstract

The scaling-up of large language models (LLMs) necessitates computing systems to have multi-processor-chip architectures, elevating the importance of chip-to-chip (C2C) communication. However, designing efficient C2C hardware architectures for LLM workloads faces three key challenges: generating realistic LLM-specific C2C traffic, accurately simulating hardware-level communication at scale, and efficiently exploring the exponentially large C2C design space. We propose C2C-Explorer, an adaptive Bayesian DSE framework that integrates a LLM-workload-driven traffic generator, a scalable interconnect simulator (switch/full-mesh, up to 512 chips), and a metric-guided evaluator into a workload-to-hardware optimization pipeline, enabling systematic C2C architectural co-design under realistic LLM workloads. Validated against FPGA-based C2C prototypes, the C2C simulator achieves 2.46-8.23% end-to-end timing error across diverse traffic patterns. Its hybrid cycle and event model further accelerates large-scale simulation by up to 7.8$\times$ over a pure cycle-accurate baseline. Applied to a 32-XPU DeepSeek-R1-671B inference workload, C2C-Explorer identifies configurations that improve goodput by 44.1% and reduce memory by 98.4%. C2C-Explorer is open-source and available at https://github.com/Selinaee/C2C-Explorer.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 11 canonical work pages

  1. [33]

    Eth-x scale up interconnect protocol test report

    Open Data Center Committee (ODCC). Eth-x scale up interconnect protocol test report. Technical Re- port ODCC-2025-03005, Open Data Center Commit- tee (ODCC), September 2025. In Chinese

    work page 2025

  2. [32]

    Eth-x scale up interconnectprotocolwhitepaper,version1.0

    Open Data Center Committee (ODCC). Eth-x scale up interconnectprotocolwhitepaper,version1.0. Techni- calReportODCC-2025-03002,OpenDataCenterCom- mittee (ODCC), September 2025. In Chinese

    work page 2025

  3. [1]

    Architecture, Chip, and Package Co-design Flow for 2.5D IC Design Enabling Heterogeneous IP Reuse

    Jinwoo Kim, Gauthaman Murali, Heechun Park, Eric Qin, Hyoukjun Kwon, Venkata Chaitanya, Krishna Chekuri, Nihar Dasari, Arvind Singh, Minah Lee, HakkiMertTorun,KallolRoy,MadhavanSwaminathan, Saibal Mukhopadhyay, Tushar Krishna, and Sung Kyu Lim. Architecture, Chip, and Package Co-design Flow for 2.5D IC Design Enabling Heterogeneous IP Reuse. In Proceeding...

    work page 2019

  4. [2]

    Ub-mesh: ahi- erarchically localized nd-fullmesh datacenter network 8 architecture

    Heng Liao, Bingyang Liu, Xianping Chen, Zhigang Guo,ChuanningCheng,JianbingWang,XiangyuChen, PengDong,RuiMeng,WenjieLiu,etal. Ub-mesh: ahi- erarchically localized nd-fullmesh datacenter network 8 architecture. arXiv preprint arXiv:2503.20377, 2025

    arXiv 2025

  5. [5]

    Thorough characterization and analysis of large transformer model training at-scale

    Scott Cheng, Jun-Liang Lin, Murali Emani, Siddhisan- ket Raskar, Sam Foreman, Zhen Xie, Venkatram Vish- wanath, and Mahmut Taylan Kandemir. Thorough characterization and analysis of large transformer model training at-scale. 8(1), February 2024. doi: 10.1145/3639034. URL https://doi.org/10.1145/ 3639034

    doi:10.1145/3639034 2024

  6. [6]

    Lang Xu, Kaushik Kandadi Suresh, Quentin Anthony, Nawras Alnaasan, and Dhabaleswar K. Panda. Char- acterizingcommunicationpatternsindistributedlarge languagemodelinference,2025. URL https://arxiv. org/abs/2507.14392

    arXiv 2025

  7. [7]

    Megatron-lm: Training multi-billion parameter language models using model parallelism, 2020

    Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catan- zaro. Megatron-lm: Training multi-billion parameter language models using model parallelism, 2020. URL https://arxiv.org/abs/1909.08053

    arXiv 2020

  8. [8]

    Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Dis- covery & Data Minin, KDD 2020, page 3505–3506. Association for Computing Machinery, 2020. ISBN 978145037998...

    arXiv 2020

Show all 36 references
  1. [9]

    Nccl: Nvidia collective commu- nications library

    NVIDIA Corporation. Nccl: Nvidia collective commu- nications library. https://developer.nvidia.com/ nccl, 2025. Accessed: 2025-11-18

    2025

  2. [11]

    Efficient Bayesian Yield Op- timization Approach for Analog and SRAM Circuits

    Mengshuo Wang, Fan Yang, Changhao Yan, Xuan Zeng, and Xiangdong Hu. Efficient Bayesian Yield Op- timization Approach for Analog and SRAM Circuits. In Proceedings of the 54th Annual Design Automation Conference, DAC 2017, pages 11:1–11:6. ACM, 2017. doi: 10.1145/3061639.3062234

    2017

  3. [12]

    Vidur: A large-scale simulation framework for llm inference

    Amey Agrawal, Nitin Kedia, Jayashree Mohan, Ashish Panwar, Nipun Kwatra, Bhargav S Gulavani, Ra- machandran Ramjee, and Alexey Tumanov. Vidur: A large-scale simulation framework for llm inference. In Proceedings of Machine Learning and Systems, 2024, 6:351–366

    2024

  4. [13]

    SimAI: Unifying architecture design and performance tuning for Large-Scale large lan- guagemodeltrainingwithscalabilityandprecision

    Xizheng Wang, Qingxu Li, Yichi Xu, Gang Lu, Dan Li, Li Chen, Heyang Zhou, Linkang Zheng, Sen Zhang, Yikai Zhu, Yang Liu, Pengcheng Zhang, Kun Qian, Kunling He, Jiaqi Gao, Ennan Zhai, Dennis Cai, and Binzhang Fu. SimAI: Unifying architecture design and performance tuning for La...

    2025

  5. [14]

    vtrain: A simulation framework for evaluating cost-effective and compute- optimal large language model training

    Jehyeon Bang, Yujeong Choi, Myeongwoo Kim, Yongdeok Kim, and Minsoo Rhu. vtrain: A simulation framework for evaluating cost-effective and compute- optimal large language model training. In 2024 57th IEEE/ACM International Symposium on Microarchitec- ture, MICRO 2024, pages 153...

    2024

  6. [15]

    Llmservingsim: A hw/sw co-simulation infrastructure for llm inference serving at scale

    Jaehong Cho, Minsu Kim, Hyunmin Choi, Guseul Heo, and Jongse Park. Llmservingsim: A hw/sw co-simulation infrastructure for llm inference serving at scale. In 2024 IEEE International Symposium on Workload Characterization, IISWC 2024, page 15–29. IEEE, September 2024. doi: 10.1...

    2024

  7. [16]

    Calculon: a methodology and tool for high-level co-design of systems and large language models

    Mikhail Isaev, Nic Mcdonald, Larry Dennison, and Richard Vuduc. Calculon: a methodology and tool for high-level co-design of systems and large language models. In Proceedings of the International Confer- ence for High Performance Computing, Networking, Storage and Analysis, SC...

    2023

  8. [17]

    A hardware evaluation framework for large language model inference, 2023

    Hengrui Zhang, August Ning, Rohan Prabhakar, and David Wentzlaff. A hardware evaluation framework for large language model inference, 2023. URL https: //arxiv.org/abs/2312.03134

    2023 arXiv

  9. [18]

    Becker, George Michelogian- nakis, James Balfour, Brian Towles, D

    Nan Jiang, Daniel U. Becker, George Michelogian- nakis, James Balfour, Brian Towles, D. E. Shaw, John Kim, and William J. Dally. A detailed and flexible cycle-accurate network-on-chip simulator. In 2013 9 IEEE International Symposium on Performance Analy- sis of Systems and So...

    2013

  10. [19]

    Kite: Afamilyofheterogeneousin- terposertopologiesenabledviaaccurateinterconnect modeling

    Srikant Bharadwaj, Jieming Yin, Bradford Beckmann, andTusharKrishna. Kite: Afamilyofheterogeneousin- terposertopologiesenabledviaaccurateinterconnect modeling. In 2020 57th ACM/IEEE Design Automation Conference, DAC 2020, pages 1–6. ACM, 2020. doi: 10.1109/DAC18072.2020.9218539

    2020

  11. [20]

    Reinhardt, Ali Saidi, Arkaprava Basu, Joel Hestness, Derek R

    Nathan Binkert, Bradford Beckmann, Gabriel Black, Steven K. Reinhardt, Ali Saidi, Arkaprava Basu, Joel Hestness, Derek R. Hower, Tushar Krishna, Somayeh Sardashti, Rathijit Sen, Korey Sewell, Muhammad Shoaib, Nilay Vaish, Mark D. Hill, and David A. Wood. The gem5 simulator. SI...

    2011

  12. [21]

    Evaluating chiplet-based Large-Scale intercon- nection networks via Cycle-Accurate Packet-Parallel simulation

    Yinxiao Feng, Yuchen Wei, Dong Xiang, and Kaisheng Ma. Evaluating chiplet-based Large-Scale intercon- nection networks via Cycle-Accurate Packet-Parallel simulation. In 2024 USENIX Annual Technical Conference, USENIX ATC 2024), pages 731–747. USENIX Association, July 2024. ISB...

    2024

  13. [22]

    Hex- amesh: Scaling to hundreds of chiplets with an op- timized chiplet arrangement

    Patrick Iff, Maciej Besta, Matheus Cavalcante, Tim Fischer, Luca Benini, and Torsten Hoefler. Hex- amesh: Scaling to hundreds of chiplets with an op- timized chiplet arrangement. In Proceedings of the 60th ACM/IEEE Design Automation Conference, DAC 2023, pages 1–6, 2023. doi: ...

    2023

  14. [23]

    T. R. Henderson, M. Lacage, et al. ns-3: A discrete- event network simulator for internet systems. In Pro- ceedings of the ACM SIGCOMM 2008 Conference on Data Communication, 2008

    2008

  15. [24]

    An overview of the omnet++ simulation environment

    András Varga and Rudolf Hornig. An overview of the omnet++ simulation environment. In Proceed- ings of the 1st International Conference on Simula- tion Tools and Techniques for Communications, Net- works and Systems & Workshops, Simutools’08. ICST (Institute for Computer Scien...

    2008

  16. [25]

    Chiplets: How Small is Too Small? In Proceedings of the 60th ACM/IEEE Design Automation Conference, DAC2023,pages1–6,2023

    AlexanderGraening,SaptadeepPal,andPuneetGupta. Chiplets: How Small is Too Small? In Proceedings of the 60th ACM/IEEE Design Automation Conference, DAC2023,pages1–6,2023. doi: 10.1109/DAC56929. 2023.10247947

    2023

  17. [26]

    Chiplet Actuary: A Quantitative Cost Model and Multi-Chiplet Archi- tecture Exploration

    Yinxiao Feng and Kaisheng Ma. Chiplet Actuary: A Quantitative Cost Model and Multi-Chiplet Archi- tecture Exploration. In Proceedings of the 59th ACM/IEEE Design Automation Conference, DAC 2022, pages 121–126. ACM, 2022. doi: 10.1145/3489517. 3530428

    2022 doi

  18. [27]

    Tracing the meta-level: Pypy’s trac- ing jit compiler

    Carl Friedrich Bolz, Antonio Cuni, Maciej Fijalkowski, and Armin Rigo. Tracing the meta-level: Pypy’s trac- ing jit compiler. ICOOOLPS 2009, page 18–25. As- sociation for Computing Machinery, 2009. ISBN 9781605585413. doi: 10.1145/1565824.1565827. URL https://doi.org/10.1145/1...

    2009

  19. [28]

    AnEffi- cientMulti-fidelityBayesianOptimizationApproachfor Analog Circuit Synthesis

    Shuhan Zhang, Wenlong Lyu, Fan Yang, Changhao Yan,DianZhou,XuanZeng,andXiangdongHu. AnEffi- cientMulti-fidelityBayesianOptimizationApproachfor Analog Circuit Synthesis. In Proceedings of the 56th Annual Design Automation Conference 2019, DAC 2019, page 64. ACM, 2019. doi: 10.1...

    2019 doi

  20. [29]

    Hanbin Hu, Peng Li, and Jianhua Z. Huang. Enabling High-Dimensional Bayesian Optimization for Efficient Failure Detection of Analog and Mixed-Signal Circuits. In Proceedings of the 56th Annual Design Automation Conference, DAC 2019, page 17. ACM. doi: 10.1145/ 3316781.3317818

    2019

  21. [30]

    Multi-objective Bayesian optimiza- tion for analog/RF circuit synthesis

    Wenlong Lyu, Fan Yang, Changhao Yan, Dian Zhou, and Xuan Zeng. Multi-objective Bayesian optimiza- tion for analog/RF circuit synthesis. In Proceedings of the 55th Annual Design Automation Conference, DAC 2018, pages 11:1–11:6. ACM, 2018. doi: 10.1145/ 3195970.3196078

    2018

  22. [31]

    An efficient Bayesian yield estimation methodforhighdimensionalandhighsigmaSRAMcir- cuits

    Jinyuan Zhai, Changhao Yan, Sheng-Guo Wang, and Dian Zhou. An efficient Bayesian yield estimation methodforhighdimensionalandhighsigmaSRAMcir- cuits. In Proceedings of the 55th Annual Design Au- tomation Conference, DAC 2018, pages 132:1–132:6. ACM, 2018. doi: 10.1145/3195970.3195987

    2018

  23. [34]

    Loop-awareoptimizationsinpypy’stracing jit

    Håkan Ardö, Carl Friedrich Bolz, and Maciej Fi- jaBkowski. Loop-awareoptimizationsinpypy’stracing jit. 48(2):63–72, October 2012. ISSN 0362-1340. doi: 10.1145/2480360.2384586

    2012

  24. [35]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shi- rong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, 10 Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu,BeiFeng,Chengd...

    2025

  25. [36]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, An- thony Hartshorn, Aobo Yang, Archi Mitra, Archie Sra- vankumar, Artem Korenev,...

    2024 arXiv

  26. [37]

    Qwen3 technical report, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chen- gen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jian- wei Zhang, Jianxin Yang,...

    2025 arXiv

  27. [96]

    doi: 10.1109/ISPASS.2013.6557149

    IEEE, 2013. doi: 10.1109/ISPASS.2013.6557149

    2013

  28. [2019]

    doi: 10.1145/3316781.3317775

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.