Pith. sign in

REVIEW 3 major objections 2 minor

MXsim is a MATLAB library that simulates OCP MX block floating-point formats with full control over exponents, precisions, and accumulator width.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-15 02:21 UTC pith:7JXVJDYO

load-bearing objection Useful MATLAB front-end for OCP MX on CPFloat; the need is real, but the abstract alone leaves the fidelity claim untested. the 3 major comments →

arxiv 2607.12915 v1 pith:7JXVJDYO submitted 2026-07-14 cs.MS cs.NAmath.NA

Simulation of Custom-Precision OCP MX Block Floating-Point Formats and Arithmetic

Maliha Islam , Mantas Mikaitis This is my paper
classification cs.MS cs.NAmath.NA
keywords OCP MXblock floating-pointmixed-precision arithmeticcustom-precision simulationMATLABCPFloatdot product accumulatorGPU formats
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces MXsim v0.1, a MATLAB library that lets researchers simulate OCP MX block floating-point arithmetic without needing the latest NVIDIA or AMD GPUs. Built on the existing custom-precision simulator CPFloat, it exposes full user control over exponents, input and output precisions, and the precision of the accumulator inside the dot product. Because GPU vendors have not published the bit-wise numerical behavior of these formats and the OCP specification itself does not prescribe precision or rounding, the library lets algorithm developers try different plausible settings and measure their effect on mixed-precision numerical methods. A sympathetic reader cares because MX formats are already shipping in hardware, yet the lack of an open, flexible software model has made it hard to design and test algorithms that will run correctly and accurately on them.

Core claim

MXsim v0.1 is a MATLAB library, layered on CPFloat, that correctly simulates OCP MX block floating-point formats and arithmetic while giving the user complete control over exponents, input/output precisions, and accumulator precision in the dot product, thereby enabling experimentation with custom MX variants and with undocumented vendor settings.

What carries the argument

MXsim itself: a MATLAB wrapper around CPFloat that implements the OCP MX block-floating-point model, exposing configurable shared exponents, element formats, and a user-chosen accumulator precision for the dot product.

Load-bearing premise

That a MATLAB simulation built on CPFloat can reproduce the numerically relevant behavior of real OCP MX hardware well enough for algorithm development, even though the exact bit-wise rules used by vendors remain unpublished.

What would settle it

Compare MXsim results for a set of mixed-precision matrix multiplications or iterative solvers against the same kernels run on actual NVIDIA or AMD MX hardware; any systematic difference in rounding error or final accuracy that cannot be explained by a documented parameter would falsify the claim of faithful simulation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Researchers can design and test mixed-precision algorithms that target MX formats without waiting for GPU access.
  • Custom MX variants (different shared-exponent widths, element precisions, or accumulator sizes) can be explored before hardware is fixed.
  • Undocumented vendor choices for rounding and intermediate precision can be reverse-engineered by matching simulation parameters to observed hardware results.
  • The same interface can be used to study the numerical stability of existing algorithms when they are moved to block floating-point.
  • Algorithm papers can report reproducible MX results that others can re-run solely in MATLAB.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Once public, MXsim could become a de-facto reference for comparing proposed MX extensions or alternative block-floating schemes.
  • The ability to vary accumulator precision independently may expose hidden sensitivity of popular mixed-precision methods to intermediate rounding that current hardware hides.
  • Porting the same control surface to other languages or to a hardware description language would let the same experiments drive silicon design decisions.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The manuscript (available here only as an abstract) presents MXsim v0.1, a MATLAB library layered on CPFloat for simulating OCP MX block floating-point formats and arithmetic. It claims user control over shared exponents, input/output element precisions, and accumulator precision in the MX-style dot product, so that researchers can experiment with custom MX variants and with undocumented vendor settings when NVIDIA/AMD GPU hardware is unavailable, and can study effects on mixed-precision algorithms given that OCP does not prescribe precision and rounding.

Significance. If the full library correctly implements the claimed customizations and is shown to capture the numerically relevant behavior of OCP MX blocks, it would be a practical open tool for the mixed-precision algorithm community. Building on CPFloat, exposing a simple MATLAB interface, and targeting the underspecification of OCP MX and vendor bit-wise behavior are genuine strengths for a software-tool contribution—provided the body supplies design detail, validation, and reproducible examples that the abstract alone cannot establish.

major comments (3)
  1. [Abstract] Abstract: The central claim that MXsim correctly simulates OCP MX block formats and arithmetic is untestable from the abstract. The full manuscript must specify how shared block exponents are extracted and applied across a block, how the custom-precision accumulator is realized inside the MX-style dot product, and how rounding modes and subnormals are handled relative to the (underspecified) OCP rules; without that design, the fidelity claim cannot be assessed.
  2. [Abstract] Abstract: No validation against known MX cases, hardware, or other simulators is mentioned. For a cs.MS software-tool paper the claim that the library is suitable for mixed-precision algorithm development is load-bearing and requires reproducible numerical examples, error checks, and at least limited comparison; their absence from the reviewed material leaves the main claim unsubstantiated.
  3. [Abstract] Abstract: The asserted “full customisation” of exponents, I/O precisions, and accumulator precision needs concrete API documentation, default choices, and correctness arguments in the body. The abstract states the capability but does not show that the configuration surface is sufficient or that it interacts correctly with CPFloat; that gap is material to the research-utility claim.
minor comments (2)
  1. [Abstract] The abstract could more clearly separate what the OCP MX specification prescribes from what MXsim chooses as defaults or free parameters, so readers know which knobs are research extensions versus compliance.
  2. [Abstract] Availability (repository, license, version pinning of CPFloat) should be stated explicitly in the full manuscript if not already present; v0.1 alone is insufficient for reproducibility.

Circularity Check

0 steps flagged

No circularity: software-tool announcement with no derivation chain, fitted predictions, or load-bearing self-citation reductions.

full rationale

Only the abstract is available. It announces MXsim v0.1, a MATLAB library layered on CPFloat for simulating OCP MX block floating-point formats and arithmetic, with user control over exponents, I/O precisions, and accumulator precision. There are no equations, no claimed first-principles derivations, no fitted parameters presented as predictions, and no uniqueness theorems or ansatzes imported via self-citation. The text is a tool description and motivation statement (undocumented vendor bit-wise behavior; OCP specification does not prescribe precision/rounding). Under the hard rules, circularity requires a quotable reduction of a claimed result to its inputs by construction; none exists here. Residual dependence on CPFloat correctness and on the authors' reading of the OCP MX specification is ordinary software-engineering risk, not circularity. Score 0 with empty steps is the correct outcome for an abstract-only software announcement.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

As a tool paper described only by abstract, the claim rests on domain assumptions about the OCP MX specification and on the correctness of the prior CPFloat simulator, plus the unstated assumption that MATLAB-level simulation of configurable block formats is a useful proxy for undocumented hardware. No free parameters are fitted; no new physical entities are postulated.

axioms (3)
  • domain assumption The OCP MX block floating-point format family is a meaningful target for mixed-precision algorithm research and is only partially specified regarding precision and rounding.
    Abstract motivation: precision and rounding are not prescribed by the OCP specification and vendor bit-wise behavior is undocumented.
  • domain assumption CPFloat correctly implements custom non-block IEEE-like floating-point formats and is a sound base for building block floating-point simulation.
    Abstract states MXsim is built on top of CPFloat for custom variants of non-block IEEE 754 formats.
  • ad hoc to paper Configurable simulation of exponents, element precisions, and accumulator precision is sufficient to explore algorithmically relevant MX variants in MATLAB.
    This is the design premise of MXsim as described; not independently validated in the abstract.
invented entities (1)
  • MXsim v0.1 MATLAB library no independent evidence
    purpose: Provide a software simulator of OCP MX block floating-point arithmetic and formats with user-configurable parameters for mixed-precision algorithm experiments.
    The library is the paper's contribution. independent_evidence is false in this review because no public code, tests, or hardware cross-checks are present in the abstract-only material.

pith-pipeline@v1.1.0-grok45 · 6106 in / 2481 out tokens · 30012 ms · 2026-07-15T02:21:50.827283+00:00 · methodology

0 comments
read the original abstract

We present MXsim v0.1: a MATLAB library for simulating the OCP MX block floating-point arithmetic and formats. The library is built on top of a custom-precision floating-point simulator, CPFloat, which allows custom variants of non-block IEEE 754 floating-point formats to be simulated. MXsim is a MATLAB simulator, and is targeted at mixed-precision algorithm development community to perform experimentation in OCP MX block floating-point arithmetic by using a simple and well-documented interface in MATLAB, which is especially useful when access to the latest NVIDIA and AMD GPUs is not available. Furthermore, due to full customisation of exponents, precisions of input and output formats, and the precision of the accumulator within the dot product operation, MXsim allows experimenting with custom variants of the OCP MX block floating point for research. Finally, since the bit-wise numerical behavior of OCP MX formats is not documented by GPU vendors, and the precision and rounding are not prescribed by the OCP specification itself, MXsim allows experiments with various possible settings in order to determine the effects on mixed-precision numerical algorithms.

discussion (0)

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