Taking a Docking Benchmark with a Known Answer onto Two GPU Servers from Different Vendors

Linxira Bio SDK cloud validation report · 2026-10-02. Environment, machine performance, and commands fully disclosed; reproduction scripts ship with the repository.

The short version

1. Docking search budget has a clear “enough” line: on the official Vina tutorial system (HIV-1 protease + indinavir), the default exhaustiveness of 8 lost the correct pose once in three seeds (off by 12 Å); at 32 and above, all seeds landed within 0.07 Å of the official pose. Use 8 for fast screening, 32 for anything you report.
2. The analysis chain passed a signal-injection test: with 150 differential genes and one deliberately enriched pathway planted in a simulated transcriptome, the five-step chain recovered the pathway at p = 6×10⁻³¹ with everything else at background level — no misses, no false alarms.
3. The two servers produced byte-identical outputs: docking, SSR mining, the expression chain, enrichment — five classes of output files have identical SHA256 hashes on both machines.
4. The Rust implementation holds a steady 3.6–4.0× over the traditional Python tool (1 M-base SSR mining, re-measured on three different CPUs), with byte-identical output to the reference implementation.
5. The domestic card runs our stack: the same kernel source compiled first-try under Moore Threads’ mcc with correct results, and its memory bandwidth utilization is in the same band as the NVIDIA card.

1 Who ran this, and why

We did — the Linxira Bio SDK project. The SDK is about to gain GPU vendor backends (including Moore Threads MUSA), and before shipping that we needed one question answered: on different vendors’ hardware, is our toolchain still correct, still fast, and still reproducible? We rented two cloud servers for a full validation round; this post is the record.

One principle up front: every exam question is an existing community benchmark, nothing home-made —

• docking uses the official AutoDock Vina tutorial system (HIV-1 protease
  • indinavir, PDB entry 1iep [3]), whose answer key ships with the tutorial (best affinity -13.234 kcal/mol and the reference pose file);
• memory bandwidth uses the STREAM convention [5];
• SSR mining uses pytrf, the community reference implementation, as the correctness control;
• transcriptomics has no single answer key, so we used signal injection — plant known signals in simulated data and have the tool find them — the standard way to validate an analysis pipeline.

Established benchmarks mean the results are directly comparable with the community, with no question of a test tailored to ourselves.

2 Environment and machine performance

Two cloud-rented servers, same-day window, same repository clone built on-site (47.3 s for 59 crates):

(Measured memory/disk figures reflect these two cloud instances under host scheduling, not vendor specifications; see §8.)

3 What was proven

3.1 Docking: correct, and with a known budget line

The redocking check — take the experimentally solved complex apart and let the program place the ligand back from a random start [2,4]. Twelve (budget × seed) combinations on each machine; the table shows heavy-atom deviation of the best pose from the official reference pose, in Å:

Both machines produced this table value-for-value identical (failed seed included). Within 1–2 Å means “back in the same pocket”; 12 Å is the classic Monte Carlo under-budget trap — no error message, just a quietly suboptimal answer (-10.9 still looks “fine”). Threads stop helping past 8 (72 s → 13 s, then flat) because e=8 is exactly 8 independent chains.

3.2 The transcriptome chain passed signal injection

A simulated profile (2000 genes × 12 samples, Poisson counts) with 150 planted differential genes and one planted enriched pathway, run through the five-step chain (QC → median-ratio normalization [6] → PCA → clustering → GO hypergeometric enrichment [7]) entirely via our Rust CLI:

• the planted pathway came out first at p = 6.3×10⁻³¹ (44 genes); all 25 others sat above p = 0.03 — nothing false surfaced;
• PCA and clustering separated the samples cleanly by condition.

3.3 Reproducibility: five output classes, byte-identical

Docking JSONs, SSR TSVs (1 M bases), same-seed docking reruns, expression chain JSONs, enrichment JSONs — SHA256-identical across the two servers. Different vendor, different CPU, different core count, same bytes.

4 Our speed against the traditional tool

SSR (microsatellite) mining is a direct same-algorithm comparison: our Rust implementation vs pytrf, the community’s standard Python implementation (which doubles as the correctness reference). 1 M-base input, median of three runs on three different CPUs:

Output files are byte-identical to the reference implementation — the speedup does not come at the cost of changed results. Worth stating about docking: our role there is orchestration and reproduction (we call the native Vina engine [2,4]); we claim no credit on the docking algorithm itself — this round proves our orchestration adds no bias.

5 The two GPUs, side by side

One kernel source, each vendor’s compiler, four kernels (copy, triad, histogram, correlation), all through their correctness gates (bit-level or < 10⁻⁴ vs an f64 reference):

Conclusion: same bandwidth league (both ~90% of spec), equivalent for data-movement-bound bioinformatics kernels; the atomics gap is worth coding around on MUSA (bucketed reductions instead of atomic counters).

6 Two driver-layer facts found along the way

Neutral records, log evidence in the repository ledgers: NVIDIA’s vGPU cloud driver ships no Vulkan entry point (three independent instances, same verbatim error); Moore Threads’ Vulkan driver loads and opens the device but fails during initialization (suspected userspace toolkit 3.1.0 vs host driver 2.7.0 pairing; evidence being packaged for vendor feedback). Neither affects this post’s results — the kernels went through each vendor’s native compiler.

7 How to reproduce

git clone https://github.com/Linxira-OS/linxira-bio-sdk.git
cd linxira-bio-sdk && cargo build --release -p linxira-bio-cli

bash scripts/rental/dock-study.sh                          # docking study + RMSD
mcc -O3 scripts/rental/bench_suite.mu -o bench_mu -lmusart # GPU kernels (MUSA)
cp scripts/rental/bench_suite.mu b.cu && nvcc -O3 -arch=sm_89 b.cu -o bench_cu
bash scripts/rental/rna_analysis.sh                        # signal-injection check
bash scripts/rental/system_bench.sh                        # storage/RAM bandwidth

Raw outputs (JSON/TSV/logs/hashes) live in the local archive release-artifacts-rental/ (not committed); per-round ledgers in docs/engine-evals/.

8 Boundaries

• One rented instance per card, one window; cloud vCPUs, memory bandwidth, and disks are subject to host scheduling — measured latencies describe these two instances, not vendor specifications.
• Disk reads are page-cache figures (containers may not drop caches); the docking conclusions hold for the official 1iep system — validate others independently.
• The e=8 failure mode is an inherent property of Monte Carlo search budget, not a defect; its value is reminding you to set your own budget line.

References

[1] Williams, S., Waterman, A., Patterson, D. Roofline: an insightful visual performance model for multicore architectures. Commun. ACM 52(4), 65–76 (2009). doi:10.1145/1498685.1498715

[2] Trott, O., Olson, A. J. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 31(2), 455–461 (2010). doi:10.1002/jcc.21334

[3] Berman, H. M. et al. The Protein Data Bank. Nucleic Acids Res. 28(1), 235–242 (2000). doi:10.1093/nar/28.1.235

[4] Eberhardt, J., Santos-Martins, D., Tillack, A. F., Forli, S. AutoDock Vina 1.2.0: New docking methods, expanded force field, and Python bindings. Nucleic Acids Res. 49(W1), W3355–W3362 (2021). doi:10.1093/nar/gkab663

[5] McCalpin, J. D. STREAM: Sustainable Memory Bandwidth in High Performance Computers. Technical Report, University of Virginia (1991–2007). https://www.cs.virginia.edu/stream/

[6] Anders, S., Huber, W. Differential expression analysis for sequence count data. Genome Biol. 11, R106 (2010). doi:10.1186/gb-2010-11-10-r106

[7] Ashburner, M. et al. Gene Ontology: tool for the unification of biology. Nat. Genet. 25(1), 25–29 (2000). doi:10.1038/75556

Version & Declarations

• Version: initial release 2026-10-02.
• Self-assessment: an author-reported validation, not third-party verified; one instance per card, one window; boundaries in §8. Third-party replication is planned follow-up.
• License: code AGPL-3.0-or-later; this article CC-BY-4.0.
• Author & provenance: Linxira-OS project maintainer · Repository: https://github.com/Linxira-OS/linxira-bio-sdk

Linxira-OS · AGPL-3.0-or-later (code) · this article CC-BY-4.0 · https://github.com/Linxira-OS/linxira-bio-sdk