Optimization Debt, Not Hardware Limits: Vector Search on IBM POWER

Cross-architecture benchmarks implicitly assume equal software optimization—an assumption that rarely holds. We quantify this effect through vector database performance on IBM POWER and AMD EPYC. In same-generation comparison (POWER10 vs EPYC 7313P, both 7nm/2021), the x86 advantage is 27% in pgvector (neither platform SIMD-optimized) but 186% in MariaDB (mature AVX-512 paths, minimal POWER VSX investment). This 7x variation suggests that over 80% of the observed gap in optimized workloads may reflect software investment asymmetry rather than architectural limitations. Compiler experiments reinforce this hypothesis: IBM XL versus GCC yields up to 8x throughput difference on identical POWER hardware. Database engine selection produces 2.1–5.5x effects—exceeding typical hardware generational improvements. These findings suggest that cross-architecture comparisons may conflate hardware capability with optimization maturity. For POWER platforms, the path to competitive vector search performance likely runs through software investment rather than hardware replacement.