pith. machine review for the scientific record. sign in

arxiv: 2604.11109 · v1 · submitted 2026-04-13 · 💻 cs.DC · cs.AI· cs.LG· cs.PF

Recognition: unknown

Record-Remix-Replay: Hierarchical GPU Kernel Optimization using Evolutionary Search

Daniel Nichols , Konstantinos Parasyris , Caetano Melone , Tal Ben-Nun , Giorgis Georgakoudis , Harshitha Menon
Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:52 UTC · model grok-4.3

classification 💻 cs.DC cs.AIcs.LGcs.PF
keywords GPU kernel optimizationevolutionary searchrecord-replay compilationhierarchical optimizationcompiler pass orderingBayesian optimizationLLM-driven searchhigh-performance computing
0
0 comments X

The pith

Record-Remix-Replay optimizes full scientific GPU applications more effectively and nearly an order of magnitude faster than traditional or modern evolutionary search methods.

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

This paper presents Record-Remix-Replay as a hierarchical framework to optimize GPU kernels in scientific computing. It combines LLM-driven evolutionary search to explore implementation choices, Bayesian optimization for parameters, and record-replay compilation to speed up evaluations. The goal is to search the entire space from source code changes through compiler flags to launch configurations without needing as much manual intervention. If successful, this would reduce the months developers spend tuning apps for new hardware generations. The approach is shown to deliver better performance on full applications while searching much quicker than isolated methods or pure evolutionary search.

Core claim

Record-Remix-Replay (R^3) enables end-to-end optimization of GPU kernels by hierarchically applying evolutionary search guided by large language models, Bayesian optimization, and record-replay techniques. This integration allows efficient exploration of a combined space including source-level changes, compiler pass sequences, and runtime parameters. As a result, it achieves superior performance on complete scientific applications compared to optimizing kernel parameters or compiler flags alone, and does so nearly ten times faster than current evolutionary search approaches.

What carries the argument

The Record-Remix-Replay (R^3) hierarchical framework, which records compilation and execution traces to enable rapid remixing and replay of optimized variants during search that spans source implementation, compiler passes, and kernel launch parameters.

Load-bearing premise

The hierarchical combination of LLM-driven evolutionary search, Bayesian optimization, and record-replay compilation can explore the full GPU optimization space scalably without adding prohibitive overhead or requiring extra human input.

What would settle it

Running Record-Remix-Replay on a full scientific application and measuring that the achieved performance is no better than tuning kernel parameters and compiler flags separately, or that total search time is not substantially lower than existing evolutionary search methods.

Figures

Figures reproduced from arXiv: 2604.11109 by Caetano Melone, Daniel Nichols, Giorgis Georgakoudis, Harshitha Menon, Konstantinos Parasyris, Tal Ben-Nun.

Figure 1
Figure 1. Figure 1: Overview of MAP-Elites evolution as in AlphaEvolve/OpenEvolve. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our Record-Remix-Replay framework. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The breakdown in times spent in generation, compiling, and evaluation [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of OpenEvolve’s LLM selection algorithm to R [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of kernel speedups across the approaches and applications [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of final achieved speedup versus the time-to-solution. The [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of kernel speedups across the approaches and applications [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of absolute times from kernels in the miniWeather [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
read the original abstract

As high-performance computing and AI workloads become increasingly dependent on GPUs, maintaining high performance across rapidly evolving hardware generations has become a major challenge. Developers often spend months tuning scientific applications to fully exploit new architectures, navigating a complex optimization space that spans algorithm design, source implementation, compiler flags and pass sequences, and kernel launch parameters. Existing approaches can effectively search parts of this space in isolation, such as launch configurations or compiler settings, but optimizing across the full space still requires substantial human expertise and iterative manual effort. In this paper, we present Record-Remix-Replay (R^3), a hierarchical optimization framework that combines LLM-driven evolutionary search, Bayesian optimization, and record-replay compilation techniques to efficiently explore GPU kernel optimizations from source-level implementation choices down to compiler pass ordering and runtime configuration. By making candidate evaluation fast and scalable, our approach enables practical end-to-end search over optimization dimensions that are typically treated separately. We show that Record-Remix-Replay can optimize full scientific applications better than traditional approaches over kernel parameters and compiler flags, while also being nearly an order of magnitude faster than modern evolutionary search approaches.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Record-Remix-Replay (R^3) hierarchical optimization framework for GPU kernels. It combines LLM-driven evolutionary search, Bayesian optimization, and record-replay compilation to explore optimizations across source-level implementation, compiler flags and passes, and kernel launch parameters. The central claim is that this approach optimizes full scientific applications better than traditional methods limited to kernel parameters and compiler flags, while achieving nearly an order of magnitude faster optimization times compared to modern evolutionary search techniques by making candidate evaluations fast and scalable.

Significance. If the results hold, this work could be significant for the field of high-performance computing and GPU-accelerated scientific applications. By hierarchically integrating established techniques with LLM-driven search, it promises to automate what currently requires substantial manual expertise, potentially accelerating the adaptation of codes to new GPU architectures. The claimed reduction in optimization time would be particularly valuable for large-scale applications where tuning is a bottleneck.

major comments (2)
  1. [Abstract] The abstract asserts concrete performance gains and speedups over traditional approaches and modern evolutionary search, but provides no experimental setup, baselines, workloads, quantitative results, tables, or error analysis to support these claims. Without this evidence, the central claims cannot be evaluated.
  2. [Framework Description] The description of the R^3 framework does not include any analysis or measurements of the overhead introduced by LLM calls, the number of LLM inferences per generation, Bayesian optimization costs, or record-replay compilation times. This information is necessary to substantiate the claim that the approach is nearly an order of magnitude faster, as unaccounted latency in these components could negate the speedup.
minor comments (2)
  1. [Abstract] The title uses 'Evolutionary Search' but the abstract emphasizes LLM-driven evolutionary search; clarifying the role of LLMs versus traditional evolutionary algorithms would improve precision.
  2. [Abstract] The term 'record-replay compilation techniques' is introduced without a brief definition or reference, which may confuse readers unfamiliar with the specific method.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. The comments highlight opportunities to improve clarity and substantiation of our claims, and we outline specific revisions below to address them.

read point-by-point responses

  1. Referee: [Abstract] The abstract asserts concrete performance gains and speedups over traditional approaches and modern evolutionary search, but provides no experimental setup, baselines, workloads, quantitative results, tables, or error analysis to support these claims. Without this evidence, the central claims cannot be evaluated.

    Authors: We appreciate this observation. The abstract is designed as a concise high-level summary of the contributions and findings, with the full experimental setup (including baselines, workloads such as the scientific applications tested, quantitative results, tables, and error analysis) presented in detail in Sections 4 and 5. To better support the claims at the abstract level, we will revise the abstract to include a brief reference to the key workloads evaluated and the magnitude of observed speedups, while directing readers to the relevant sections for complete details. revision: yes

  2. Referee: [Framework Description] The description of the R^3 framework does not include any analysis or measurements of the overhead introduced by LLM calls, the number of LLM inferences per generation, Bayesian optimization costs, or record-replay compilation times. This information is necessary to substantiate the claim that the approach is nearly an order of magnitude faster, as unaccounted latency in these components could negate the speedup.

    Authors: We agree that a breakdown of these overhead components is important for rigorously supporting the speedup claims. The manuscript currently focuses on the hierarchical integration and end-to-end results. We will add a new subsection (likely in Section 3) that reports measurements of LLM inference overheads, the number of inferences per generation, Bayesian optimization costs, and record-replay compilation times, including how these contribute to the overall nearly order-of-magnitude improvement relative to standard evolutionary search. revision: yes

Circularity Check

0 steps flagged

No circularity: framework combines established techniques with no derivations or self-referential reductions

full rationale

The paper describes Record-Remix-Replay (R^3) as a hierarchical framework integrating LLM-driven evolutionary search, Bayesian optimization, and record-replay compilation to explore GPU kernel optimizations. No equations, first-principles derivations, fitted parameters presented as predictions, or load-bearing self-citations appear in the provided text. Claims rest on empirical evaluation of the combined system rather than any step that reduces by construction to its inputs. This is a standard engineering contribution presenting a new composition of prior methods, with no evidence of self-definitional, fitted-input, or uniqueness-imported circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the premise that the described hierarchical search strategy is both effective and scalable for real scientific workloads. Because only the abstract is available, the ledger records the introduction of the R^3 framework itself as the primary addition beyond prior techniques.

axioms (1)
  • domain assumption GPU kernel optimization involves a complex space spanning algorithm design, source implementation, compiler flags and pass sequences, and kernel launch parameters.
    Explicitly stated in the abstract as the space developers must navigate.
invented entities (1)
  • Record-Remix-Replay (R^3) hierarchical optimization framework no independent evidence
    purpose: To combine LLM-driven evolutionary search, Bayesian optimization, and record-replay compilation for end-to-end GPU kernel tuning.
    New method introduced by the paper to address the stated optimization challenge.

pith-pipeline@v0.9.0 · 5525 in / 1491 out tokens · 46127 ms · 2026-05-10T15:52:35.616880+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages · 5 internal anchors

  1. [1]

    Opentuner: an extensible framework for program autotuning,

    J. Ansel, S. Kamil, K. Veeramachaneni, J. Ragan-Kelley, J. Bosboom, U.-M. O’Reilly, and S. Amarasinghe, “Opentuner: An extensible framework for program autotuning,” in2014 23rd International Conference on Parallel Architecture and Compilation Techniques (PACT), 2014, pp. 303–315. [Online]. Available: https://doi.org/10. 1145/2628071.2628092

    work page arXiv 2014

  2. [2]

    Active harmony: Towards automated performance tuning,

    C. Tapus, I.-H. Chung, and J. Hollingsworth, “Active harmony: Towards automated performance tuning,” inSC ’02: Proceedings of the 2002 ACM/IEEE Conference on Supercomputing, 2002, pp. 44–44. [Online]. Available: https://doi.org/10.1109/SC.2002.10062

    work page doi:10.1109/sc.2002.10062 2002

  3. [3]

    Kernel launcher: C++ library for optimal-performance portable cuda applications,

    S. Heldens and B. van Werkhoven, “Kernel launcher: C++ library for optimal-performance portable cuda applications,”arXiv preprint arXiv:2303.12374, 2023. [Online]. Available: https://doi.org/10.48550/ arXiv.2303.12374

    work page arXiv 2023

  4. [4]

    Cltune: A generic auto-tuner for opencl kernels,

    C. Nugteren and V . Codreanu, “Cltune: A generic auto-tuner for opencl kernels,” in2015 IEEE 9th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC). Los Alamitos, CA, USA: IEEE Computer Society, sep 2015, pp. 195–202. [Online]. Available: https://doi.org/10.1109/MCSoC.2015.10

    work page doi:10.1109/mcsoc.2015.10 2015

  5. [5]

    Piecewise holistic autotuning of compiler and runtime parameters,

    M. Popov, C. Akel, W. Jalby, and P. de Oliveira Castro, “Piecewise holistic autotuning of compiler and runtime parameters,” inEuro- Par 2016: Parallel Processing, P.-F. Dutot and D. Trystram, Eds. Cham: Springer International Publishing, 2016, pp. 238–250. [Online]. Available: https://doi.org/10.1007/978-3-319-43659-3 18

    work page doi:10.1007/978-3-319-43659-3 2016

  6. [6]

    CERE: llvm-based codelet extractor and replayer for piecewise benchmarking and optimization,

    P. de Oliveira Castro, C. Akel, E. Petit, M. Popov, and W. Jalby, “CERE: llvm-based codelet extractor and replayer for piecewise benchmarking and optimization,”ACM Trans. Archit. Code Optim., vol. 12, no. 1, pp. 6:1–6:24, 2015. [Online]. Available: https://doi.org/10.1145/2724717

    work page doi:10.1145/2724717 2015

  7. [7]

    Merchandiser: Data placement on heterogeneous memory for task-parallel hpc applications with load- balance awareness,

    Z. Xie, J. Liu, J. Li, and D. Li, “Merchandiser: Data placement on heterogeneous memory for task-parallel hpc applications with load- balance awareness,” inProceedings of the 28th ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming, ser. PPoPP ’23. New York, NY , USA: Association for Computing Machinery, 2023, p. 204–217. [Onli...

    work page arXiv 2023

  8. [8]

    Scalable Tuning of (OpenMP) GPU Applications via Kernel Record and Replay,

    K. Parasyris, G. Georgakoudis, E. Rangel, I. Laguna, and J. Doerfert, “Scalable Tuning of (OpenMP) GPU Applications via Kernel Record and Replay,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, ser. SC ’23. New York, NY , USA: Association for Computing Machinery, Nov. 2023, pp. 1–14. [Online...

    work page doi:10.1145/3581784.3607098 2023

  9. [9]

    A Survey of Evolutionary Algorithms,

    L. Liu, T. Fei, Z. Zhu, K. Wu, and Y . Zhang, “A Survey of Evolutionary Algorithms,” in2023 4th International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE), Aug. 2023, pp. 22–27. [Online]. Available: https://doi.org/ 10.1109/ICBAIE59714.2023.10281260

    work page doi:10.1109/icbaie59714.2023.10281260 2023

  10. [10]

    Mathematical discoveries from program search with large language models,

    B. Romera-Paredes, M. Barekatain, A. Novikov, M. Balog, M. P. Kumar, E. Dupont, F. J. R. Ruiz, J. S. Ellenberg, P. Wang, O. Fawzi, P. Kohli, and A. Fawzi, “Mathematical discoveries from program search with large language models,”Nature, vol. 625, no. 7995, pp. 468–475, Jan

  11. [11]

    , author Barekatain, M

    [Online]. Available: https://doi.org/10.1038/s41586-023-06924-6

    work page doi:10.1038/s41586-023-06924-6

  12. [12]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery

    A. Novikov, N. V ˜u, M. Eisenberger, E. Dupont, P.-S. Huang, A. Z. Wagner, S. Shirobokov, B. Kozlovskii, F. J. R. Ruiz, A. Mehrabian, M. P. Kumar, A. See, S. Chaudhuri, G. Holland, A. Davies, S. Nowozin, P. Kohli, and M. Balog, “AlphaEvolve: A coding agent for scientific and algorithmic discovery,” Jun. 2025. [Online]. Available: https://doi.org/10.48550/...

    work page internal anchor Pith review doi:10.48550/arxiv.2506.13131 2025

  13. [13]

    Bar- barians at the gate: How ai is upending systems research.arXiv 13 Georgios Liargkovas, Mihir Nitin Joshi, Hubertus Franke, and Kostis Kaffes preprint arXiv:2510.06189, 2025

    A. Cheng, S. Liu, M. Pan, Z. Li, B. Wang, A. Krentsel, T. Xia, M. Cemri, J. Park, S. Yang, J. Chen, L. Agrawal, A. Desai, J. Xing, K. Sen, M. Zaharia, and I. Stoica, “Barbarians at the Gate: How AI is Upending Systems Research,” Oct. 2025. [Online]. Available: https://doi.org/10.48550/arXiv.2510.06189

    work page doi:10.48550/arxiv.2510.06189 2025

  14. [14]

    Openevolve: an open-source evolutionary cod- ing agent,

    A. Sharma, “Openevolve: an open-source evolutionary cod- ing agent,” 2025. [Online]. Available: https://github.com/ algorithmicsuperintelligence/openevolve

    2025

  15. [15]

    control bars

    J.-B. Mouret and J. Clune, “Illuminating search spaces by mapping elites,” Apr. 2015. [Online]. Available: https://doi.org/10.48550/arXiv. 1504.04909

    work page internal anchor Pith review doi:10.48550/arxiv 2015

  16. [16]

    Kernel tuner: A search-optimizing gpu code auto- tuner,

    B. van Werkhoven, “Kernel tuner: A search-optimizing gpu code auto- tuner,”Future Generation Computer Systems, vol. 90, pp. 347–358,

  17. [17]

    Available: https://doi.org/10.1016/j.future.2018.08.004

    [Online]. Available: https://doi.org/10.1016/j.future.2018.08.004

    work page doi:10.1016/j.future.2018.08.004 2018

  18. [18]

    TVM: An Automated End-to-End Optimizing Compiler for Deep Learning,

    T. Chen, T. Moreau, Z. Jiang, L. Zheng, E. Yan, M. Cowan, H. Shen, L. Wang, Y . Hu, L. Ceze, C. Guestrin, and A. Krishnamurthy, “TVM: An Automated End-to-End Optimizing Compiler for Deep Learning,” Feb

  19. [19]

    Available: https://doi.org/10.48550/arXiv.1802.04799

    [Online]. Available: https://doi.org/10.48550/arXiv.1802.04799

    work page doi:10.48550/arxiv.1802.04799

  20. [20]

    Ansor: Generating high-performance tensor programs for deep learning,

    L. Zheng, C. Jia, M. Sun, Z. Wu, C. H. Yu, A. Haj-Ali, Y . Wang, J. Yang, D. Zhuo, K. Sen, J. E. Gonzalez, and I. Stoica, “Ansor: Generating high-performance tensor programs for deep learning,” arXiv, 2020, oSDI 2020 (arXiv version). [Online]. Available: https://doi.org/10.48550/arXiv.2006.06762

    work page doi:10.48550/arxiv.2006.06762 2020

  21. [21]

    Tensor program optimization with probabilistic programs,

    J. Shao, X. Zhou, S. Feng, B. Hou, R. Lai, H. Jin, W. Lin, M. Masuda, C. H. Yu, and T. Chen, “Tensor program optimization with probabilistic programs,” arXiv, 2022, accepted to NeurIPS 2022 (arXiv version). [Online]. Available: https://doi.org/10.48550/arXiv.2205.13603

    work page doi:10.48550/arxiv.2205.13603 2022

  22. [22]

    Learning to optimize halide with tree search and random programs,

    A. Adams, K. Ma, L. Anderson, R. Baghdadi, T.-M. Li, M. Gharbi, B. Steiner, S. Johnson, K. Fatahalian, F. Durand, and J. Ragan-Kelley, “Learning to optimize halide with tree search and random programs,” ACM Transactions on Graphics, vol. 38, no. 4, pp. 121:1–121:12,

  23. [23]

    Available: https://doi.org/10.1145/3306346.3322967

    [Online]. Available: https://doi.org/10.1145/3306346.3322967

    work page doi:10.1145/3306346.3322967

  24. [24]

    Triton: an intermediate language and compiler for tiled neural network computations,

    P. Tillet, H. T. Kung, and D. Cox, “Triton: An intermediate language and compiler for tiled neural network computations,” inProceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages (MAPL 2019), 2019, pp. 10–19. [Online]. Available: https://doi.org/10.1145/3315508.3329973

    work page doi:10.1145/3315508.3329973 2019

  25. [25]

    Micomp: Mitigating the compiler phase-ordering problem using optimization sub-sequences and machine learning,

    A. H. Ashouri, A. Bignoli, G. Palermo, C. Silvano, S. Kulkarni, and J. Cavazos, “Micomp: Mitigating the compiler phase-ordering problem using optimization sub-sequences and machine learning,”ACM Trans. Archit. Code Optim., vol. 14, no. 3, Sep. 2017. [Online]. Available: https://doi.org/10.1145/3124452

    work page doi:10.1145/3124452 2017

  26. [26]

    Compilergym: Robust, performant compiler optimization environments for ai research,

    C. Cummins, B. Wasti, J. Guo, B. Cui, J. Ansel, S. Gomez, S. Jain, J. Liu, O. Teytaud, B. Steiner, Y . Tian, and H. Leather, “Compilergym: Robust, performant compiler optimization environments for ai research,” in2022 IEEE/ACM International Symposium on Code Generation and Optimization (CGO), 2022, pp. 92–105. [Online]. Available: https://doi.org/10.1109/...

    work page doi:10.1109/cgo53902.2022.9741258 2022

  27. [27]

    Apollo: Reusable models for fast, dynamic tuning of input-dependent code,

    D. Beckingsale, O. Pearce, I. Laguna, and T. Gamblin, “Apollo: Reusable models for fast, dynamic tuning of input-dependent code,” in2017 IEEE International Parallel and Distributed Processing Symposium (IPDPS). IEEE, 2017, pp. 307–316. [Online]. Available: https://doi.org/10.1109/IPDPS.2017.38

    work page doi:10.1109/ipdps.2017.38 2017

  28. [28]

    Artemis: Automatic runtime tuning of parallel execution parameters using machine learning,

    C. Wood, G. Georgakoudis, D. Beckingsale, D. Poliakoff, A. Gimenez, K. Huck, A. Malony, and T. Gamblin, “Artemis: Automatic runtime tuning of parallel execution parameters using machine learning,” inHigh Performance Computing, B. L. Chamberlain, A.-L. Varbanescu, H. Ltaief, and P. Luszczek, Eds. Cham: Springer International Publishing, 2021, pp. 453–472. ...

    work page doi:10.1007/978-3-030-78713-4 2021

  29. [29]

    Auto-tuning parameter choices in hpc applications using bayesian optimization,

    H. Menon, A. Bhatele, and T. Gamblin, “Auto-tuning parameter choices in hpc applications using bayesian optimization,” in2020 IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2020, pp. 831–840. [Online]. Available: https://doi.org/10. 1109/IPDPS47924.2020.00090

    work page arXiv 2020

  30. [30]

    Gptune: Multitask learning for autotuning exascale applications,

    Y . Liu, W. M. Sid-Lakhdar, O. Marques, X. Zhu, C. Meng, J. W. Demmel, and X. S. Li, “Gptune: Multitask learning for autotuning exascale applications,” inProceedings of the 26th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, ser. PPoPP ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 234–246. [Online]. Avai...

    work page doi:10.1145/3437801.3441621 2021

  31. [31]

    An architecture for flexible auto-tuning: The periscope tuning framework 2.0,

    R. Mijakovi ´c, M. Firbach, and M. Gerndt, “An architecture for flexible auto-tuning: The periscope tuning framework 2.0,” in 2016 2nd International Conference on Green High Performance Computing (ICGHPC), 2016, pp. 1–9. [Online]. Available: https: //doi.org/10.1109/ICGHPC.2016.7508066

    work page doi:10.1109/icghpc.2016.7508066 2016

  32. [32]

    Proteus: Portable Runtime Optimization of GPU Kernel Execution with Just-in-Time Compilation,

    G. Georgakoudis, K. Parasyris, and D. Beckingsale, “Proteus: Portable Runtime Optimization of GPU Kernel Execution with Just-in-Time Compilation,” inProceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization, ser. CGO ’25. New York, NY , USA: Association for Computing Machinery, Mar. 2025, pp. 507–522. [Online]. Available...

    work page doi:10.1145/3696443.3708939 2025

  33. [33]

    Integrating performance tools in model reasoning for gpu kernel optimization,

    D. Nichols, K. Parasyris, C. Jekel, A. Bhatele, and H. Menon, “Integrating performance tools in model reasoning for gpu kernel optimization,” 2025, https://doi.org/10.48550/arXiv.2510.17158. [Online]. Available: https://arxiv.org/abs/2510.17158

    work page doi:10.48550/arxiv.2510.17158 2025

  34. [34]

    Agentic ai security: Threats, defenses, evaluation, and open challenges,

    D. Nichols, P. Polasam, H. Menon, A. Marathe, T. Gamblin, and A. Bhatele, “ Performance-Aligned LLMs for Generating Fast HPC Code ,”IEEE Transactions on Parallel & Distributed Systems, no. 01, pp. 1–12, Mar. 5555, https://doi.org/10.1109/TPDS.2026.3675550. [Online]. Available: https://doi.ieeecomputersociety.org/10.1109/TPDS. 2026.3675550

    work page doi:10.1109/tpds.2026.3675550 2026

  35. [35]

    Efficient Memory Management for Large Language Model Serving with PagedAttention

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica, “Efficient Memory Management for Large Language Model Serving with PagedAttention,” Sep. 2023, arXiv:2309.06180 [cs]. [Online]. Available: https://doi.org/10.48550/ arXiv.2309.06180

    work page internal anchor Pith review arXiv 2023

  36. [36]

    Lulesh 2.0 updates and changes,

    I. Karlin, J. Keasler, and J. R. Neely, “Lulesh 2.0 updates and changes,” Lawrence Livermore National Laboratory (LLNL), Tech. Rep., 07

  37. [37]

    Available: https://doi.org/10.2172/1090032

    [Online]. Available: https://doi.org/10.2172/1090032

    work page doi:10.2172/1090032

  38. [38]

    Assessing a mini-application as a performance proxy for a finite element method engineering application,

    P. T. Lin, M. A. Heroux, R. F. Barrett, and A. B. Williams, “Assessing a mini-application as a performance proxy for a finite element method engineering application,”Concurrency and Computation: Practice and Experience, vol. 27, no. 17, pp. 5374–5389, 2015. [Online]. Available: https://doi.org/10.1002/cpe.3587

    work page doi:10.1002/cpe.3587 2015

  39. [39]

    The scalable heterogeneous computing (shoc) benchmark suite,

    A. Danalis, G. Marin, C. McCurdy, J. S. Meredith, P. C. Roth, K. Spafford, V . Tipparaju, and J. S. Vetter, “The scalable heterogeneous computing (shoc) benchmark suite,” inProceedings of the 3rd Workshop on General-Purpose Computation on Graphics Processing Units, ser. GPGPU-3. New York, NY , USA: Association for Computing Machinery, 2010, p. 63–74. [Onl...

    work page doi:10.1145/1735688.1735702 2010

  40. [40]

    miniweather,

    M. R. Norman, “miniweather,” [Computer Software], March 2020. [Online]. Available: https://doi.org/10.11578/dc.20201001.88

    work page doi:10.11578/dc.20201001.88 2020

  41. [41]

    arXiv: 2303.14006 ISBN: 9798350397390

    Z. Jin and J. S. Vetter, “A Benchmark Suite for Improving Performance Portability of the SYCL Programming Model,” in2023 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), Apr. 2023, pp. 325–327. [Online]. Available: https://doi.org/10.1109/ISPASS57527.2023.00041

    work page doi:10.1109/ispass57527.2023.00041 2023

  42. [42]

    2546–2554

    J. Bergstra, R. Bardenet, Y . Bengio, and B. K ´egl, “Algorithms for hyper-parameter optimization,” inProceedings of the 25th International Conference on Neural Information Processing Systems, ser. NIPS’11. Red Hook, NY , USA: Curran Associates Inc., Dec. 2011, pp. 2546–2554, https://dl.acm.org/doi/10.5555/2986459.2986743. [Online]. Available: https://pap...

    work page doi:10.5555/2986459.2986743 2011

  43. [43]

    gpt-oss-120b & gpt-oss-20b Model Card

    OpenAI, S. Agarwal, L. Ahmad, J. Ai, S. Altman, A. Applebaum, E. Arbus, R. K. Arora, Y . Bai, B. Baker, H. Bao, B. Barak, A. Bennett, T. Bertao, N. Brett, E. Brevdo, G. Brockman, S. Bubeck, C. Chang, K. Chen, M. Chen, E. Cheung, A. Clark, D. Cook, M. Dukhan, C. Dvorak, K. Fives, V . Fomenko, T. Garipov, K. Georgiev, M. Glaese, T. Gogineni, A. Goucher, L. ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.10925 2025

  44. [44]

    OpenAI GPT-5 System Card

    A. Singh, A. Fry, A. Perelman, A. Tart, A. Ganesh, A. El-Kishky, A. McLaughlin, A. Low, A. J. Ostrow, A. Ananthram, A. Nathan, A. Luo, A. Helyar, A. Madry, A. Efremov, A. Spyra, A. Baker- Whitcomb, A. Beutel, A. Karpenko, A. Makelov, A. Neitz, A. Wei, A. Barr, A. Kirchmeyer, A. Ivanov, A. Christakis, A. Gillespie, A. Tam, A. Bennett, A. Wan, A. Huang, A. ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.03267 2025

  45. [45]

    Solving lattice qcd systems of equations using mixed precision solvers on gpus,

    M. Clark, R. Babich, K. Barros, R. Brower, and C. Rebbi, “Solving lattice qcd systems of equations using mixed precision solvers on gpus,” Computer Physics Communications, vol. 181, no. 9, pp. 1517–1528,

  46. [46]

    Available: https://doi.org/10.1016/j.cpc.2010.05.002

    [Online]. Available: https://doi.org/10.1016/j.cpc.2010.05.002

    work page doi:10.1016/j.cpc.2010.05.002 2010