Disclosed and Undisclosed GPU Hardware Exclusion Across the Modern Protein Structure Prediction and Design Toolchain: A Compute-Accessibility Audit

Generative and predictive protein structure tools increasingly rely on bfloat16-optimized, Triton-compiled kernels for performance, a dependency that silently requires Ampere-generation or newer NVIDIA GPUs (compute capability ≥ 8.0) and excludes the Turing-generation hardware (compute capability 7.0-7.5) still common in academic compute clusters. While benchmarking generative protein-binder-design tools against amyloid and tau fibril targets, we independently and repeatedly encountered this barrier with one specific tool (ESMFold2), which fails outright on Turing hardware with no documentation warning users in advance. Motivated by this, we conducted a systematic compute-accessibility audit of 13 major protein structure prediction and design tools released between 2021 and 2026, assessing each tool's actual hardware requirements against primary sources (official documentation, source code, and GitHub issue trackers) and, where we had direct access, empirical testing on Turing-generation hardware. We find that hardware exclusion is not predicted by a tool's age: RFdiffusion3 (2025) runs on Turing GPUs with no modification, while AlphaFold3 and Chai-1 (both 2024) do not. More strikingly, tools that fail on older hardware do so in qualitatively different and unequally documented ways. We propose a five-category taxonomy of outcomes: no exclusion; disclosed hard exclusion with a clean failure (Chai-1); disclosed hard exclusion with silent data corruption rather than an error (AlphaFold3, confirmed via the developers' own bug-tracker responses); undisclosed hard exclusion with a clean failure (ESMFold2); and undisclosed exclusion with an available graceful-degradation fallback (Boltz-1/Boltz-2). We additionally document a symmetric, opposite-direction failure mode in which a tool (the original RFdiffusion) breaks on GPUs newer than its intended target due to stale CUDA/PyTorch version pinning. We argue that undocumented hardware requirements, and particularly silent-corruption failures, represent an underappreciated and unequally distributed barrier to computational biology research, and we recommend that tool developers disclose minimum compute capability explicitly alongside every release.