## Goals

- **Be the fastest GPU columnar data processing library on AMD hardware.** Match or exceed cuDF/RAPIDS performance on equivalent NVIDIA hardware using Vulkan compute shaders — prove that vendor-neutral APIs don't sacrifice performance.

- **Run everywhere.** A single shared library (`libvulkan_columnar.so` / `.dll`) that executes on AMD RDNA/CDNA, NVIDIA (via Vulkan), Intel Arc, Apple Silicon (MoltenVK), and a software fallback (lavapipe). No CUDA, no ROCm, no HIP — just Vulkan 1.3.

- **Integrate into anything.** A C ABI that lets Polars, DataFusion, Spark, DuckDB, and Python all call the same GPU primitives. The library doesn't pick a DataFrame — it empowers all of them.

- **Correctness you can bet on.** Every shader tested against a CPU reference. Every operation benchmarked. Every PR validated with ASan, UBSan, and Vulkan validation layers. No silent data corruption, ever.

- **Batch-first, single-GPU, visible win.** Designed for throughput workloads (gigabytes per kernel launch). No clusters required to see the speedup. No streaming micro-batches that pay kernel launch tax. If it ships, it's measurably faster than CPU on one consumer GPU.