I spent months building an exact incremental-computation library in C. I finally put the reproducible benchmarks on GitHub.

I’ve been working on a performance project called HKD Kernel.

The premise is that a lot of large computations repeat work unnecessarily. If only a tiny portion of the input changed, you ideally want to recompute only the state affected by that change rather than rebuild everything.

HKD Kernel is a native C implementation of that idea with exact-result checks.

The most surprising result for me has been the size of the gap on highly sparse workloads. Across the benchmark configuration currently documented in the repository, the measured mean speedup is around 18,000x versus the full-recomputation path.

Huge caveat: that is not an “all software becomes 18,000x faster” claim. The gain comes specifically from workloads where most previously computed state remains valid.

I made the benchmark reproducible because I’d rather have people try to break the claim than just believe a graph.

GitHub:

https://github.com/yangofzeal/hkd-kernel

I’m particularly looking for:

adversarial benchmark ideas
real workloads with sparse state changes
feedback on the C API
cases where you think incremental recomputation would be the wrong architecture

There’s a free/community path and I’m also testing a $999/month production-node Business license, but right now I’m much more interested in whether developers can reproduce or invalidate the technical results.