Answers are grounded in this paper and cite their sources.
The associative data protection mechanism hides masked bits of reference patterns (etalons) in pseudorandom filler, expanding data thirtyfold; recovery matches sections against them. We evaluate the performance, not the security, of offloading it to graphics processing units (GPUs) in .NET (ILGPU); the implementation stores its key with the data and must not be used in security-sensitive settings. On two laptops with integrated (OpenCL) and discrete (CUDA) GPUs, the file-based full cycle broke even with the library’s sequential central processing unit (CPU) implementation at 10–100 KB and 10–30 KB and was 4.3–9.2 times faster at 100 MB. Against an optimized (vectorized, table-driven) CPU path, neither GPU was faster from 1 KB to 100 MB on the second laptop. With reused buffers, coalesced stores, and overlapped file operations, the discrete GPU was not distinguishable from this path at 100 MB (time ratio 0.97 [0.90, 1.04], which does not establish equivalence) and 1.1 times slower at 10 MB. By approximate estimates obtained under differing measurement conditions, kernels took 5–6% of formation and at most 22% of recovery on the discrete GPU. Hardware counters attribute its slow search kernel to thread divergence (1.6 of 32 threads active per warp instruction on uniform data), yet a 26–37-fold faster table-driven kernel did not detectably change the full cycle. For dichotomic key masks, recognition provably reduces to a table lookup (at most 512 entries), which the library uses to validate keys. The results characterize one implementation and do not establish an advantage of GPUs for this transform.
No takes yet. Share an insight, caveat, or question.
Gibadullin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: