HorneSci

Reduce token consumption and energy use by eliminating redundant calculations and reusing what’s already computed.

Simulations, solvers, and monitoring systems often apply the same matrix, grid, or network to an input that changes a little at a time, and still recompute the whole result every step. We change that code to update the result from the part of the input that changed, and we measure the new version against the old one on your hardware.

Delta Routing and HST

HorneSci develops two complementary technologies: Delta Routing and HST. Delta Routing acts like a GPS for computation, intelligently selecting the most efficient method for each task. HST is one of those methods, designed to accelerate certain sparse-matrix operations. Together, they aim to reduce unnecessary computation, improve processing speed, and lower energy costs.

A computation arrives at Delta Routing, which chooses one of four methods for each call: HST sparse, HST dense, a column delta, or full recompute. Each path returns the same result. Incoming computation A fixed matrix and a few changed inputs Delta Routing Evaluates the computation and chooses an efficient method one method per call HST sparse Prepared tiles HST dense Packed dense delta Column delta Changed columns only Full recompute When most inputs change Efficient computation Only the work the change requires A computation arrives at Delta Routing, which chooses one of four methods for each call: HST sparse, HST dense, a column delta, or full recompute. Each path returns the same result. Incoming computation A fixed matrix and a few changed inputs Delta Routing Evaluates the computation and chooses an efficient method one method per call HST sparse Prepared tiles HST dense Packed dense delta Column delta Changed columns only Full recompute When most inputs change Efficient computation Only the work the change requires
Delta Routing checks each workload and sends every call down one path. The two HST methods are shown in bold outline.

Where the work repeats

Simulation

Simulations that revisit the same system while only a small region of the grid changes.

Rows of consoles in Mission Control, Houston, in 1965, with flight controllers at work.

Monitoring

Warehouse sensor events normalized into one update stream.

A long row of supercomputer racks with perforated doors, seen at an angle.

Data centers

Rollups from sensor to rack to room, rebuilt from the leaves when a few readings move.

Fixed cameras

Video analysis where most of each frame matches the frame before.

Contact us

  1. Call

    30 minutes · Available 24/7

    Want your servers to stop recomputing what hasn’t changed? Talk with HorneSci about how we’ll get there.

  2. Assessment

    2–4 weeks, fixed fee

    We run your workload on your machines and report what we measured, the conditions behind each number, and the changes that would produce it.

  3. Deployment

    In your codebase

    We make the change inside your program, in its own language, with your engineers.

Book a call

We work under your NDA and on your machines, connected or air-gapped. Your data stays with you, and we publish only what you approve in writing.

A 1963 computer room: tape drives along the wall, a woman loading a reel, a printer and an operator in the foreground.

Delta Routing

Delta Routing, our public project, updates the output of a fixed operator from the part of its input that changed.

Updates only the outputs a change reaches.

Results match the full product to 1e-9, and sampled audits check the routed path against it.

More about Delta Routing at deltarouting.com

About

HorneSci is a software and research company with two cofounders.

About HorneSci
Portrait of Hunter Horne.

Hunter Horne

Cofounder, mathematics and theory

LinkedIn
Portrait of Eric Spencer.

Eric Spencer

Cofounder, implementation

LinkedIn

Member of the NVIDIA Inception program and AWS Activate.

  • HorneSci is a member of the NVIDIA Inception program.
  • Powered by AWS. HorneSci is a member of AWS Activate.
HorneSci