Validation of Cross-Platform Determinism (ARM/CPU vs H100/GPU) using FPBench

[Eric Waller]

Dear FPBench Community,

This email shares recent results using the FPBench standard to verify cross-platform determinism for LuxiEdge, a high-throughput deterministic math engine.

Context & Architecture LuxiEdge is architected for industrial and edge applications where strict bit-exactness is required alongside high efficiency (e.g., regulated financial modeling or autonomous edge systems). The engine operates as a stateless vector processor, designed to maintain identical numerical outputs across disparate hardware architectures (such as ARM Neon edge devices vs. NVIDIA H100 datacenter GPUs).

Given these constraints, FPBench was selected as the "gold standard" to validate the engine’s cross-platform bit-exactness claims.

Scope of Validation To ensure transparency regarding the testing scope: This validation targeted the numerical core. The test suite filtered the FPBench library for straight-line, single-variable expressions, excluding those requiring while loops or separate state management, as the current focus was verifying the stateless math pipeline.

Methodology The validation compared outputs from two distinct environments:

1. Consumer Edge: Mac M1 Pro (ARM/Neon)
2. Datacenter AI: NVIDIA H100 (CUDA)

Parameters:

• Input: 100,000 f32 values per expression.
• Distribution: Random Uniform [0.1, 3.14] (seeded).
• Implementation: Uniform software-defined math stack (Rust/libm core) to ensure algorithmic consistency across architectures.

Results LuxiEdge achieved a 100% SHA-256 hash match on the output vectors for all 59 tested expressions across both platforms, validating that strict cross-platform determinism is achievable even at high throughputs on disparate hardware.

Expression
Mac M1 Pro Hash
H100 GPU Hash
Status
sin(x)
f0cc...6c328b3f6a6c7
f0cc...6c328b3f6a6c7
MATCH
(exp(x)-1)/x
323e...9ffce84402ea7f53
323e...9ffce84402ea7f53
MATCH
...full logs available.

Note: Independent validation of these metrics (via TestFort QA Lab) and full citation formats are available on the project website (luxiedge.com) for those interested in reproducible research data.

These findings are submitted to contribute to the community's discussion on cross-platform reproducibility in high-performance computing.

Humbly submitted,

Eric Waller 

LuxiEdge