hrxdb

An embedded GPU vector index for AMD Strix Halo, written in Rust. Loom kernels perform cosine search and top-k; HRX owns the resident buffers and executes the work. Custom kernels can use the same corpus without reading intermediate vectors back to the CPU.

Recorded workload	Median completion time
One top-10 query over 10M × 384 vectors	33.7 ms
Sixty top-5 queries over 6,909,092 × 384 vectors	80.7 ms
Append 50 rows to 7M × 768 vectors	0.54 ms
Replace 100 scattered rows in an unshared 7M × 768 corpus	0.88 ms

Local gfx1151 measurements from September 2026 with generated data; ingestion and compilation are excluded. The rows are separate workloads, not a batching speedup comparison. Results and methodology.

Features

• Exhaustive cosine search over FP16 vectors, with FP32 scores and deterministic ties.
• Up to 256 queries per batch and top-k up to 1,024.
• Append and update snapshots while existing readers keep their original rows.
• Shared corpus handles, independent search workspaces and native snapshot files.
• Device queries, row gathering and custom scoring followed by GPU top-k.

Search is exact over the stored representation; quantization can change rankings relative to the original vectors. Embeddings, metadata and server APIs belong to the application.

Quick start

Requires Rust 1.91+, Linux x86-64 and Strix Halo (gfx1151).

[dependencies]
hrxdb = "0.3.1"

use hrxdb::{Corpus, Device};

fn main() -> hrxdb::Result<()> {
    let device = Device::open(0)?;
    let corpus = Corpus::build(&device, 3, [
        [1.0, 0.0, 0.0],
        [0.0, 1.0, 0.0],
        [0.8, 0.2, 0.0],
    ])?;
    let mut searcher = corpus.searcher()?;
    let neighbors = searcher.search(&[1.0, 0.0, 0.0], 2)?;
    assert_eq!(neighbors[0].id, 0);
    Ok(())
}

IDs are insertion positions. Corpus::build normalizes and quantizes FP32 input; build_fp16 preserves supplied FP16 values. From a checkout, run cargo run --locked --release --example search.

First GPU use downloads HRX's native bundle. Execution requires amdgpu/KFD access, glibc 2.43+ and compatible system libraries. Native setup.

Use an existing HRX context

Corpus::build_in and build_fp16_in allocate in an encoder's ModelContext. Keep queries, gathered rows and scores resident through a custom pipeline, then download the selected IDs and scores.

Storage and snapshots · Custom scoring · Layouts and synchronization · API reference

Development

cargo test --locked --all-targets
cargo test --locked --doc
cargo clippy --locked --all-targets -- -D warnings
cargo run --locked --release --bin hrxdb-bench -- --rows 100000 --samples 10

CPU checks need no GPU. Hardware tests run separately:

HRX_OFFLINE=1 cargo test --locked --release -- --ignored --test-threads=1 --skip batch::bench

The largest test allocates about 9.4 GB. Profiling · Contributing · Changelog · Releases

License

MIT. The HRX runtime and compiler have separate third-party notices.