HorneSci

Compute only what changed.

Get results faster on the hardware you already use.

What we do

We speed up repeated calculations by reusing results that still hold when the input changes. We measure the difference on your hardware.

Incoming callSparse example matrixInput values are below the matrix. The output values are at its right.

Delta Routing

Matches full recompute

  1. Sparse updateChanged columns, sparse

  2. Dense updateChanged columns, dense

  3. Column deltaChanged columns only

  4. Full recomputeWhen most inputs change

Testimonials

46% less energy per job

“HST finished the job in 12.2 seconds instead of 19.8, and used 133 joules instead of 249.”

Bharat SharmaWattOps · Read the WattOps case study

“Your tech is demonstrably successful.”

Chris WebbEnergy Capital ·

What it saves

An estimate for your hardware, from our own benchmarks.

No saving above 50% changed on CPU, 32% on GPU.

Energy saved per year
—
​
Cost saved per year
—
Work skipped per unit
—
10 units
——
1,000 units
——
100,000 units
——

Estimated from HorneSci benchmarks at rated power: tuned OpenBLAS on one CPU core (28 W), cuBLAS on a GPU (350 W).

Work with us

Three ways to work together, each starting with a call.

  1. Assessment

    We run your workload on your machines and report each number with its conditions.

    2–4 weeks, fixed fee

  2. Deployment

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

    Your code, your engineers

  3. Software

    Delta Routing CLI, an Ubuntu image with Delta Routing built in, and the Delta Routing Console desktop app.

    Private evaluation

Start with a call

Thirty minutes, any hour. Under your NDA, on your machines, connected or air-gapped.

Where it applies

How the Dirty Set Moves

Direct incremental update speedup versus fastest full recomputationAcross 21 SuiteSparse matrices in 400-step sessions, direct column updates beat the fastest of eight full implementations. At 0%, 0.2%, 1%, 5%, and 25% dirty-set turnover per step, local drift speedups were 2.10, 2.29, 2.44, 3.04, and 4.02 times. Uniform jump speedups were 2.12, 2.25, 3.09, 5.85, and 8.19 times.Column updates vs. full recompute1×4×8×0%5%15%25%Dirty-set turnover per stepLocal driftUniform jump
Horne et al. · Table III

HorneSci in three minutes

An introduction to the company and how we work.

A computing room at NACA Langley in 1957: mathematicians at desks in the foreground, operators at an IBM 704 console, tape drives, and card equipment behind them.Play the film 3:07

About

We find repeated work in software, then remove it.

  • Portrait of Hunter Horne.

    Hunter Horne

    Cofounder, mathematics and theory

    LinkedIn
  • Portrait of Eric Spencer.

    Eric Spencer

    Cofounder, implementation

    LinkedIn