Skip to content
H98
H98

Simulations that get faster every time you run them.

An automated simulation workflow manager that reuses past results and speeds up the math, so every run costs less than the last.

Join a pilotJoin the waitlist

Henry Cantor, Hansen Peng, and Nathan Jeong

The problem

Simulations start cold, every single time.

Density functional theory alone uses more than 30% of the yearly compute cycles at NERSC, the Department of Energy's main supercomputing center. Most of it starts from zero.

  • Near-repeats run from scratch

    A new structure or a tweaked parameter still means recomputing work that is mostly done already.

  • Caching only catches exact copies

    Tools like AiiDA reuse results only when the entire job is identical. Change one input, reuse nothing.

  • Results are scattered and purged

    Outputs sit across scratch, local disks and archives with no index: hard to find, easy to lose.

“We must regularly remove the computed data from the /scratch directory. Once results are moved, it's inconvenient to revisit them.”
Research engineer

The solution

Two efficiency engines that multiply.

Engine 1

Mathematical Methods

Ozaki-style integer math turns high-precision matrix multiplication into exact integer products, plus other compute-level speedups.

Engine 2

WarmStart

Matches segments of a new simulation to segments of past runs, then reuses or warm-starts them, even when the overall job differs.

Result

A higher efficiency multiple

Each engine's speedup stacks on the other's, run after run.

One automated workflow manager. Every simulation result is consolidated into a single indexed database, so both engines always know what has been computed before.

Data network and results

Smarter every run, and provably so.

~1,600×
faster rerunning the aspirin PySCF script: 22 minutes down to under a second.
27×
faster on a brand-new script in fast mode: 23 minutes down to 51 seconds.
2.7×
faster on a brand-new script in exact mode, with PySCF's own answers.
0.35kcal/mol
average error in fast mode, about 0.02%.

Proof of concept

What to expect

Measured back to back on a MacBook Air (M2, 16 GB), B3LYP/def2-SVP. Exact mode returns the code's own answers; fast mode stays within a stated tolerance.

  • Run it again, or at a new temperatureExact

    ~1,600×

    The same script

    22 min0.7–0.9 s

  • Change the last step's functionalFast

    ~500×

    PBE0 or PBE final single point

    22 min3 s

  • A new functional, all of itFast

    104×

    The whole script in PBE

    22 min13 s

  • Another conformerFast

    67×

    The script, acid group turned over

    22 min20 s

  • Something brand newFast

    27×

    The aspirin script, first run

    23 min51 s

  • Quantum ESPRESSOFast

    17–20×

    Si8, atoms moved up to 0.08 Å (a full relaxation: 7×)

    85 s5 s

Metadata network

Pooled run metadata, per industry

Feeds R&D in every segment and compares wall time used vs. requested to predict the right allocation for each job.

Customers contribute run metadata only:

  • Code and version
  • Hardware, cores, memory
  • Cutoffs
  • Success, failure, error and fix
  • SCF iterations
  • Convergence curves
  • k-points, smearing, mixing

The index

Every result becomes a node in one index web.

New runClose matchesOther results

  • Indexed by calculation. Each result is catalogued by the calculation types and script sections inside it, with a similarity vector to every other result.

  • Light and confidential. Indexes and metadata live in our database. Full result files stay in your own long-term storage.

  • Built for teams. Project folders, a choice of which storage WarmStart draws from, access control and file permissions.

  • Storage that manages itself. Low-relevance runs are flagged for authorized archiving or deletion; we never delete. Pinned Artifacts are never purged.

Markets and expansion

Seven simulation markets, one reuse engine.

We start where both engines pay off, then expand segment by segment.

  1. 1

    Now

    Ozaki + WarmStart

    Bound by dense linear algebra and matrix multiplication, so integer math and warm starts both pay off. The biggest multiple.

    • EQC

      Quantum chemistry

      Electronic structure: electrons in molecules and solids (DFT, Hartree–Fock, coupled cluster).

      Packages we integrate with
      VASP, Quantum ESPRESSO, Gaussian, ORCA, PySCF
      Industry
      materials, batteries, chemicals, semiconductors
      Research
      materials, pharma, national labs
  2. 2

    Next

    WarmStart

    Memory-bandwidth bound. Caching identical runs is already standard here, proof the market wants smarter reuse.

    • FEA

      Finite element

      Packages we integrate with
      Abaqus, Ansys Mechanical, LS-DYNA, Nastran
      Buyers
      auto, aerospace, civil, consumer products
    • FLD

      Fluid dynamics

      Packages we integrate with
      Fluent, STAR-CCM+, OpenFOAM
      Buyers
      aerospace, auto, energy, turbomachinery
  3. 3

    Then

    WarmStart + new math

    Warm starts first; segment-specific multipliers added as we research them, the way Ozaki was for EQC.

    • MOL

      Molecular dynamics

      Packages we integrate with
      GROMACS, LAMMPS, AMBER, OpenMM
      Buyers
      pharma, polymers, materials
    • ELE

      Electromagnetics

      Packages we integrate with
      HFSS, CST, SPICE, Synopsys, Cadence
      Buyers
      semiconductors, telecom, defense
    • PFN

      Fusion and nuclear

      Packages we integrate with
      GENE, TRANSP, OpenMC, MCNP
      Buyers
      fusion startups, fission, national labs
  4. 4

    Pivot

    Custom optimization

    Monte Carlo and stochastic models. We custom-optimize in-house codes for banks and insurers: a pivot if needed.

    • FIN

      Financial

      Packages we integrate with
      Monte Carlo, in-house codes, lattice QCD
      Buyers
      HFTs, banks, insurers

Start with an audit

We lead with an audit of each lab's own runs. The multiple it proves goes straight into allocation proposals and grant reports.

Join a pilot

Competition

Existing tools each solve one piece.

  • Open-source math

    Ozaki libraries

    Emulate FP64 matrix math on fast integer units; now also in NVIDIA cuBLAS.

    H98: Runs still start from zero. We stack the math with reuse, and it integrates with our software.

  • Open-source workflow

    AiiDA and QCArchive

    Track provenance; reuse a job only if its input hash exactly matches a past one.

    H98: We match segments, so near-repeats get reused, not just exact copies.

  • Pay-per-use cloud

    Rowan and Mat3ra

    Cloud chemistry workflows, paid in credits per CPU or GPU minute.

    H98: A flat subscription on customers' own codes, reducing the compute needed.

  • In-house tools

    Custom pipelines

    Proprietary queueing and bookkeeping built for a single organization.

    H98: No upkeep burden, plus segment-level reuse and cross-industry learning.

Our platform: segment-level WarmStart × Ozaki-style math, running on an index of everything you've computed. Certified, and smarter with every simulation.

Team

The founders.

Get started

Put H98 on your own simulations.

  • Installs in about five minutes on Mac or Linux
  • Works with any PySCF script through one flag
  • Identical Quantum ESPRESSO, VASP, ORCA and Gaussian jobs replay; changed Quantum ESPRESSO jobs start warm
  • Calculations never leave your computer

Join a pilot

Pilots are open now in quantum chemistry. Tell us what you run and we'll set you up.

Prefer email? h98.gpus@gmail.com