Pith. sign in

REVIEW 3 major objections 2 minor 1 cited by

Inter-APU Communication on AMD MI300A Systems via Infinity Fabric: a Deep Dive

T0 review · 3 major / 2 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 This submission's abstract and body are two different papers; the body argues that reduction-based interpreters can be generated from just a syntax specification and a contraction function, while the advertised inter-APU benchmark study is

desk verdict The submission is internally mismatched: abstract promises an MI300A benchmark study; the full text is an Agda paper on reduction semantics, so the advertised work is absent. read the letter →

arxiv 2508.11298 v2 pith:UFT52CFJ submitted 2025-08-15 cs.DC

Gabin Schieffer , Jacob Wahlgren , Ruimin Shi , Edgar A. León , Roger Pearce , Maya Gokhale , Ivy Peng This is my paper · ORCID
classification cs.DC MSC 68N1868N3003B40
keywords reductionsemanticsgenericprogrammingzippersone-holecontextsparamorphismsdependenttypesAgdanormalization
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 submission is internally mismatched: the abstract promises an experimental benchmark study of inter-APU communication on AMD MI300A systems, but the supplied full text is a programming-languages paper on generic reduction-based interpreters. Only the latter has a body that can be read and checked. That paper's claim is that reduction semantics for deterministic standard strategies can be implemented without hand-written context-manipulation boilerplate, starting from only a syntax specification and a contraction function. The key insight is that evaluation contexts are zippers, and zippers are derivatives of the syntax data type, so decomposition and recomposition can be defined generically and correctly by construction. A sympathetic reader would care because this turns interpreter construction into a two-part specification problem rather than a hand-written recursive program.

What carries the argument

The central identity is that an evaluation context is a one-hole context, and for a regular data type the one-hole context is the derivative: $\partial T$ is exactly the type of zippers over $T$. Plugging a term into the hole is the generic recompose operation, and a paramorphism over the syntax supplies decompose. This carries the argument because it moves context handling from hand-written recursive functions into generic datatype algebra, making the rest of the semantics follow from the syntax description.

What would settle it

Take a deterministic standard strategy whose redex condition depends on information not present in the raw syntax type, for example well-typedness or scope safety, and try to derive decompose solely from the derivative of that syntax. If the derived decompose cannot enforce the side condition, the claim that syntax plus contraction suffices for all such strategies fails.

Watch

Extended reading notes

Core claim

For deterministic standard reduction strategies, the paper establishes that every semantic artifact beyond the syntax and the contraction function is generic: the context type, decompose, recompose, and drive can be defined once and for all in a dependently typed language, using the derivative of the syntax type as the type of evaluation contexts and a paramorphism as the decomposition function. This yields reduction-based interpreters for arithmetic expressions and for the lambda calculus with shift/reset, with no manually authored context code. Where the standard recipe treats decomposition and recomposition as interpreter-specific glue, the paper treats them as consequences of the syntax'

Load-bearing premise

The framework applies only to languages whose terms decompose uniquely into a reduction context and a redex; if one term can be decomposed two ways, the generic decompose cannot be total.

Editorial extensions

If this is right

  • Interpreter code shrinks to a syntax declaration and a contraction function; the context machinery is shared, so hand-written decompose and recompose errors disappear.
  • The construction is correct by construction: because decompose is a paramorphism and plugging comes from the derivative, one-step reduction is guaranteed to recompose what it decomposes.
  • The approach transfers across dependently typed languages such as Agda, Idris, Coq, and Lean, and can be ported to non-dependent languages by sacrificing totality guarantees.
  • Control effects such as shift/reset, which usually complicate context manipulation, are handled by the same generic machinery, as the paper demonstrates.

Reading between the lines

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

  • The derivative-as-zipper identification suggests a natural next step to multi-hole contexts through dissections, which would cover reduction strategies that need more than one hole per context step.
  • A concrete testable extension is a code generator that takes only a syntax specification plus a contraction function and emits a full normalization function; if this works for a range of calculi, the framework becomes a practical interpreter-generation pipeline.
  • Porting the framework to a non-dependent language would quantify exactly how much boilerplate returns as runtime totality checks, giving a trade-off measure for the claim that boilerplate is eliminated.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The submitted manuscript claims, in its abstract, to present a systems study of inter-APU communication on AMD MI300A systems: it promises dedicated benchmarks for direct GPU memory access, explicit inter-APU data movement, and collective multi-APU communication; a comparison of HIP, MPI, and RCCL; and optimized versions of Quicksilver and CloverLeaf running on a four-APU node. The full text supplied, however, is a different paper: Casper Bach's 'Generic Reduction-Based Interpreters' (arXiv:2508.11297v1, cs.PL), about deriving reduction-semantics interpreters from syntax specifications using generic programming in Agda. The body contains no mention of MI300A, Infinity Fabric, HIP, MPI, RCCL, Quicksilver, CloverLeaf, benchmarks, or experimental results. The central claims of the abstract are therefore unsupported by any content in the submission.

Significance. If the advertised MI300A inter-APU communication study existed, it could provide practically useful guidance for programming four-APU nodes with Infinity Fabric, particularly for the choice of programming interface, allocator, and data-movement strategy. That contribution would be within the scope of a systems venue and potentially valuable. However, as submitted, no such contribution is present in the body. The only text that could be checked is an incomplete literate-Agda draft on reduction semantics, which is unrelated to the advertised topic and does not permit any assessment of the systems claims. Thus the significance of the paper cannot be assessed from this artifact.

major comments (3)
  1. [Abstract vs. Full Text] The abstract's central claim is entirely absent from the body. The full text is a different manuscript, 'Generic Reduction-Based Interpreters' by Casper Bach, and contains no section, table, figure, or equation describing MI300A, Infinity Fabric, HIP, MPI, RCCL, Quicksilver, CloverLeaf, or any benchmark or measurement. No experimental evidence exists in this submission to support the advertised contribution.
  2. [Experimental methodology (missing)] Even if one ignored the title/body mismatch, the claimed benchmark study requires, at minimum: the platform configuration, the Infinity Fabric topology, allocator settings, the benchmark harness, the measured bandwidth/latency numbers, and the application runs for Quicksilver and CloverLeaf. None of this is present. The abstract's phrase 'four MI100A APU system' is itself inconsistent with the advertised MI300A focus and is not elaborated anywhere. The load-bearing factual content of the paper is therefore unverifiable.
  3. [Incomplete and unrelated full text] The provided full text is a partial Agda paper that ends mid-appendices and lacks sections 3–6 and parts of the reference list. As a self-contained artifact, it cannot be evaluated as a programming-languages paper either. For the advertised systems paper, this body is simply the wrong content, and the omission cannot be repaired by local revisions within this manuscript.
minor comments (2)
  1. [Abstract] The abstract refers to a 'four MI100A APU system'; given the title and surrounding text, this is presumably a typo for 'MI300A.' Please correct.
  2. [Metadata] The body's title, author, and affiliation (Casper Bach, University of Southern Denmark) do not match the systems topic implied by the abstract. If this is a submission error, the correct manuscript should be submitted; the mismatched text should not be forwarded for review.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified; supplied full text is a different manuscript, so the abstract's systems claims have no derivation to audit.

full rationale

The abstract advertises a systems paper on AMD MI300A inter-APU communication benchmarks, HIP/MPI/RCCL comparisons, and Quicksilver/CloverLeaf optimizations. The supplied full text, however, is 'Generic Reduction-Based Interpreters' by Casper Bach (arXiv:2508.11297v1, cs.PL), which contains no MI300A, Infinity Fabric, RCCL, Quicksilver, CloverLeaf, or experimental evaluation. Consequently, the claimed derivation chain of the advertised systems paper cannot be walked; there is no body text whose equations or fitted parameters could be checked for circularity. This absence of support is a manuscript/integrity issue, not a circularity pattern. For the programming-languages text actually present, the contribution is a constructive framework: evaluation contexts are represented as zippers, given by derivatives of data types (citing McBride), and decompose/recompose are derived generically from a syntax specification and a contraction function. These are standard mathematical constructions rather than predictions fitted to data or conclusions imported from the author's own prior work. The paper explicitly notes that some definitions are elided for brevity and invites readers to consult the literate Agda source; that is an omitted-detail caveat, not a circular step. No self-citation appears load-bearing, and no target result is baked into the stated assumptions. Thus, under the hard rule that circularity must be exhibited by specific reduction, no circular step can be identified in either the advertised (absent) derivation or the supplied PL derivation. Score 0.

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

The advertised systems paper cannot be audited because its body is absent; there is no measured data, so no fitted parameters can be identified. The Agda text that is present relies on standard type-theoretic machinery, but the excerpt is incomplete and the relevant assumptions are not fully stated in the supplied portion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inter-APU Communication on AMD MI300A Systems via Infinity Fabric: a Deep Dive." pith.science (2026). https://pith.science/paper/UFT52CFJ

@misc{pith2026250811298,
  author       = {Pith},
  title        = {Pith review of: Inter-APU Communication on AMD MI300A Systems via Infinity Fabric: a Deep Dive},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UFT52CFJ}},
  note         = {Machine review of arXiv:2508.11298}
}
read the original abstract

The ever-increasing compute performance of GPU accelerators drives up the need for efficient data movements within HPC applications to sustain performance. Proposed as a solution to alleviate CPU-GPU data movement, AMD MI300A Accelerated Processing Unit (APU) combines CPU, GPU, and high-bandwidth memory (HBM) within a single physical package. Leadership supercomputers, such as El Capitan, group four APUs within a single compute node, using Infinity Fabric Interconnect. In this work, we design specific benchmarks to evaluate direct memory access from the GPU, explicit inter-APU data movement, and collective multi-APU communication. We also compare the efficiency of HIP APIs, MPI routines, and the GPU-specialized RCCL library. Our results highlight key design choices for optimizing inter-APU communication on multi-APU AMD MI300A systems with Infinity Fabric, including programming interfaces, allocators, and data movement. Finally, we optimize two real HPC applications, Quicksilver and CloverLeaf, and evaluate them on a four MI100A APU system.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Enhanced Sensitivity and Noise Resilience in Two-Qubit Quantum Magnetometers

    quant-ph 2025-08 unverdicted novelty 4.0 of 10

    The abstract claims a novel two-qubit magnetometer with derived sensitivity and noise measures, but the full text is an unrelated GPU all-reduce paper, so the claim is unsupported by the submission.

Reference graph

Works this paper leans on

30 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [1]

    On Computational Small Steps and Big Steps: Refocusing for Outermost Reduction

    2015. On Computational Small Steps and Big Steps: Refocusing for Outermost Reduction. Ph. D. Dissertation. Department of Computer Science, Aarhus University

    work page 2015

  2. [2]

    Michael Gordon Abbott, Thorsten Altenkirch, Neil Ghani, and Conor McBride. 2003. Derivatives of Containers. In Typed Lambda Calculi and Applications, 6th International Conference, TLCA 2003, Valencia, Spain, June 10-12, 2003, Proceedings (Lecture Notes in Computer Science, Vol. 2701), Martin Hofmann (Ed.). Springer, 16–30. doi: 10.1007/3-540- 44904-3_2

    doi:10.1007/3-540- 2003

  3. [3]

    Andreas Abel and James Chapman. 2014. Normalization by Evaluation in the Delay Monad: A Case Study for Coinduction via Copatterns and Sized Types. In Proceedings 5th Workshop on Mathematically Structured Functional Programming, MSFP@ETAPS 2014, Grenoble, France, 12 April 2014 (EPTCS, Vol. 153), Paul Blain Levy and Neel Krishnaswami (Eds.). 51–67. doi:10.4...

    work page doi:10.4204/eptcs.153.4 2014

  4. [4]

    Mads Sig Ager, Olivier Danvy, and Jan Midtgaard. 2005. A functional correspondence between monadic evaluators and abstract machines for languages with computational effects. Theor. Comput. Sci. 342, 1 (2005), 149–172. doi:10.1016/J.TCS.2005.06.008

    work page doi:10.1016/j.tcs.2005.06.008 2005

  5. [5]

    Guillaume Allais, Robert Atkey, James Chapman, Conor McBride, and James McKinna. 2021. A type- and scope-safe universe of syntaxes with binding: their semantics and proofs. J. Funct. Program. 31 (2021), e22. doi:10.1017/S0956796820000076

    work page doi:10.1017/s0956796820000076 2021

  6. [6]

    Malgorzata Biernacka, Dariusz Biernacki, Sergueï Lenglet, and Alan Schmitt. 2022. Non-Deterministic Abstract Machines. In 33rd Inter- national Conference on Concurrency Theory, CONCUR 2022, September 12-16, 2022, Warsaw, Poland (LIPIcs, Vol. 243) , Bartek Klin, Slawomir Lasota, and Anca Muscholl (Eds.). Schloss Dagstuhl - Leibniz-Zentrum für Informatik, 7...

    work page doi:10.4230/lipics.concur.2022.7 2022

  7. [7]

    Malgorzata Biernacka, Witold Charatonik, and Klara Zielinska. 2017. Generalized Refocusing: From Hybrid Strategies to Abstract Machines. In 2nd International Conference on Formal Structures for Computation and Deduction, FSCD 2017, September 3-9, 2017, Oxford, UK (LIPIcs, Vol. 84), Dale Miller (Ed.). Schloss Dagstuhl - Leibniz-Zentrum für Informatik, 10:1...

    work page doi:10.4230/lipics.fscd.2017.10 2017

  8. [8]

    Malgorzata Biernacka and Olivier Danvy. 2007. A syntactic correspon- dence between context-sensitive calculi and abstract machines. Theor. Comput. Sci. 375, 1-3 (2007), 76–108. doi:10.1016/J.TCS.2006.12.028

    work page doi:10.1016/j.tcs.2006.12.028 2007

Show all 30 references
  1. [9]

    Venanzio Capretta. 2005. General recursion via coinductive types. Log. Methods Comput. Sci. 1, 2 (2005). doi:10.2168/LMCS-1(2:1)2005

    2005 doi

  2. [10]

    James Chapman, Pierre-Évariste Dagand, Conor McBride, and Peter Morris. 2010. The gentle art of levitation. InProceeding of the 15th ACM SIGPLAN international conference on Functional programming, ICFP 2010, Baltimore, Maryland, USA, September 27-29, 2010, Paul Hudak and Steph...

    2010

  3. [11]

    Mosses, Neil Sculthorpe, and Paolo Torrini

    Martin Churchill, Peter D. Mosses, Neil Sculthorpe, and Paolo Torrini

  4. [12]

    The Agda Community. [n. d.]. Agda Standard Library. https://github. com/agda/agda-stdlib

  5. [13]

    Nils Anders Danielsson. 2012. Operational semantics using the par- tiality monad. In ACM SIGPLAN International Conference on Func- tional Programming, ICFP’12, Copenhagen, Denmark, September 9-15, 2012, Peter Thiemann and Robby Bruce Findler (Eds.). ACM, 127–138. doi:10.1145/2...

    2012

  6. [14]

    Olivier Danvy. 2008. Defunctionalized interpreters for programming languages. In Proceeding of the 13th ACM SIGPLAN international con- ference on Functional programming, ICFP 2008, Victoria, BC, Canada, September 20-28, 2008, James Hook and Peter Thiemann (Eds.). ACM, 131–142....

    2008

  7. [15]

    Olivier Danvy. 2008. From Reduction-Based to Reduction-Free Nor- malization. In Advanced Functional Programming, 6th International School, AFP 2008, Heijen, The Netherlands, May 2008, Revised Lectures (Lecture Notes in Computer Science, Vol. 5832) , Pieter W. M. Koopman, Rinus...

    2008 doi

  8. [16]

    Olivier Danvy and Andrzej Filinski. 1990. Abstracting Control. In Proceedings of the 1990 ACM Conference on LISP and Functional Pro- gramming, LFP 1990, Nice, France, 27-29 June 1990 , Gilles Kahn (Ed.). ACM, 151–160. doi: 10.1145/91556.91622

    1990

  9. [17]

    Olivier Danvy and Jacob Johannsen. 2013. From Outermost Reduction Semantics to Abstract Machine. In Logic-Based Program Synthesis and Transformation, 23rd International Symposium, LOPSTR 2013, Madrid, Spain, September 18-19, 2013, Revised Selected Papers (Lecture Notes in Comp...

    2013 doi

  10. [18]

    Olivier Danvy, Jacob Johannsen, and Ian Zerny. 2011. A walk in the semantic park. In Proceedings of the 2011 ACM SIGPLAN Workshop on Partial Evaluation and Program Manipulation, PEPM 2011, Austin, TX, USA, January 24-25, 2011 , Siau-Cheng Khoo and Jeremy G. Siek (Eds.). ACM, 1...

    2011

  11. [19]

    Olivier Danvy and Lasse R. Nielsen. 2004. Refocusing in Reduction Semantics. BRICS Research Series RS-04-26. Department of Computer Science, Aarhus University. http://www.brics.dk/RS/04/26/

    2004

  12. [20]

    Matthias Felleisen. 1988. The Theory and Practice of First-Class Prompts. In Conference Record of the Fifteenth Annual ACM Symposium on Principles of Programming Languages, San Diego, California, USA, January 10-13, 1988 , Jeanne Ferrante and Peter Mager (Eds.). ACM Press, 180...

    1988

  13. [21]

    Matthias Felleisen and Robert Hieb. 1992. The Revised Report on the Syntactic Theories of Sequential Control and State. Theor. Comput. Sci. 103, 2 (1992), 235–271. doi:10.1016/0304-3975(92)90014-7

    1992 doi

  14. [22]

    Yannick Forster, Fabian Kunze, and Marc Roth. 2020. The weak call- by-value 𝜆-calculus is reasonable for both time and space. Proc. ACM Program. Lang. 4, POPL (2020), 27:1–27:23. doi:10.1145/3371095

    2020 doi

  15. [23]

    Gérard P. Huet. 1997. The Zipper.J. Funct. Program. 7, 5 (1997), 549–554. doi:10.1017/S0956796897002864

    1997 doi

  16. [24]

    Gilles Kahn. 1987. Natural Semantics. In STACS 87, 4th Annual Sym- posium on Theoretical Aspects of Computer Science, Passau, Germany, February 19-21, 1987, Proceedings (Lecture Notes in Computer Science, Vol. 247), Franz-Josef Brandenburg, Guy Vidal-Naquet, and Martin Wirsing...

    1987 doi

  17. [25]

    McCarthy, Jon Rafkind, Sam Tobin-Hochstadt, and Robert Bruce Findler

    Casey Klein, John Clements, Christos Dimoulas, Carl Eastlund, Matthias Felleisen, Matthew Flatt, Jay A. McCarthy, Jon Rafkind, Sam Tobin-Hochstadt, and Robert Bruce Findler. 2012. Run your research: on the effectiveness of lightweight mechanization. In Proceedings of the 39th ...

    2012

  18. [26]

    Ugo Dal Lago and Simone Martini. 2008. The weak lambda calculus as a reasonable machine. Theor. Comput. Sci. 398, 1-3 (2008), 32–50. doi:10.1016/J.TCS.2008.01.044

    2008 doi

  19. [27]

    P. J. Landin. 1964. The Mechanical Evaluation of Expressions. Comput. J. 6, 4 (1964), 308–320. doi:10.1093/COMJNL/6.4.308

    1964 doi

  20. [28]

    Conor McBride. 2001. The Derivative of a Regular Type is its Type of One-Hole Contexts. (2001)

    2001

  21. [29]

    Conor McBride. 2008. Clowns to the left of me, jokers to the right (pearl): dissecting data structures. In Proceedings of the 35th ACM Generic Reduction-Based Interpreters Conference’17, July 2017, Washington, DC, USA left-most =∀ {x xs ys frm c t 0 f c 1 t1} → S[ d∼ x ]≡ xs +...

    2008

  22. [2015]

    LNCS Trans

    Reusable Components of Semantic Specifications. LNCS Trans. Aspect Oriented Softw. Dev. 12 (2015), 132–179. doi: 10.1007/978-3-662- 46734-3_4

    2015 doi

Pith tools

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