Scale Out on Hardware You Own
Split one calculation over your own servers with a built-in job queue, reproducible sums and checkpoints. Slurm, PBS and LSF submission is in build.
2 Checked (exact)3 Partial1 Unverified1 in build, not releasedEngine 1.10.12 live
Scale Out, Step by Step
How one calculation spreads over your nodes and survives a lost one.
Space or K plays and pauses, the arrow keys step. With reduced motion every step is shown at once.
1. Pair your nodes Your own servers, paired once. Slurm, PBS and LSF submission is in build. Keys stay on your agent server.
What You Can Compute
Distributed SCF
One SCF over up to 64 ranks with a fixed summation order: same ranks, same bits.
HPC Schedulers
Slurm, PBS and LSF submission through a server you pair is in build, not released. Today the queue is built into the Chemistry Lab.
Checkpoints and Recovery
SCF checkpoints, rank-loss recovery and distributed MP2.
VYROX Math Kernel
Our own dense linear algebra and GPU integral kernels, no third-party math library.
Cluster Queue
A queue with priorities (urgent, age, fair share, size), a plain-words reason for every wait, an estimated start, fill gaps for short jobs, hold, release and cancel. Partial: tested with automated harnesses, not yet on a customer site.
Split Frequencies
Hessian displacements run in parts on several servers and merge to the same result (tested on two servers, difference 2.4e-4 cm-1).
Every Method, with Its Level of Support
Generated from the engine's own capability report (engine 1.10.12), so this list never runs ahead of the code.
Checked reproduces printed values Checked (exact) no printed value exists; checked against exact mathematics or brute force Partial code exists, unchecked parts badged in the app Unverified built, no printed check yet In build, not released in build, not released (not in the installed engine)
| Method | Support | Since | Primary Sources |
|---|---|---|---|
| Plane-wave materials: elastic constants, EOS, Bader, slabs, PDOSClamped-ion elastic constants, equation of state, Bader charges, work functions, slabs and projected DOS. Silicon lattice constant and elastic constants (clamped and relaxed c44) compared with Nielsen and Martin 1985 Table I, with a different pseudopotential, so differences of up to 8% remain (elastic constants within 6%, bulk modulus 8%). Bader charges, work functions, slabs and PDOS are not checked against printed values. | Partial | 1.9.0 | Nielsen, Martin 1985 |
| Distributed SCFOne SCF split over up to 64 ranks with bitwise reproducible sums. No paper prints a value for this. Checked by recomputation: the energy of one SCF split over 2 and 3 ranks equals the energy of the same SCF on one rank within 1e-10 Eh (Hartree-Fock, closed and open shell), repeated runs with the same rank count are identical bit for bit, and the Coulomb and exchange matrices built in pieces equal the serial ones within 1e-12 (within 1e-11 when built over three ranks). Not checked: more than 3 ranks in one run, Kohn-Sham runs, real network links (the test ranks exchange data in memory) and any speed-up. | Checked (exact) | 0.7.24 | |
| Slurm, PBS and LSF submission (scheduler bridge)Jobs sent to your own scheduler through its head node, with job-state parsers for squeue, sacct, sinfo, qstat and bjobs. In build, not released: the parsers pass their own tests, but the bridge has not run on a real scheduler site and is not installed on the portal. | In build, not released | 0.7.40 | |
| Rank-loss recovery, distributed MP2, SCF checkpointsSurvive a lost node, split MP2, resume from checkpoints. | Partial | 1.3.0 | |
| Split frequencies over serversHessian rows computed in parts on several servers and merged. No paper prints a value for this. Checked by recomputation: the nine Hessian rows of water (Hartree-Fock, STO-3G) computed in three separate parts and merged give the same frequencies and infrared intensities as one run with the same Cartesian displacements, bit for bit. Not checked: parts computed on servers with different processors, which need not round identically. A single run by default displaces along the vibrational directions instead, so it differs from a split run in the fourth decimal of a wavenumber: 7e-4 cm-1 for water on one machine, 2.4e-4 cm-1 in one measured run split over two servers (engine 1.9.8). | Checked (exact) | 0.7.40 | |
| VYROX Math KernelOur own dense linear algebra: packed GEMM, blocked Cholesky, LU and TRSM, and a threaded symmetric eigensolver, with no third-party math library. GEMM kernels for Apple SME and ARM NEON with a scalar fallback, and a blocked Householder tridiagonalisation followed by the QL algorithm. A single-thread GEMM is also used inside the meta-GGA grid code (since 1.9.8). Auto mode routes GEMM and the eigensolver only. One measured speed: dgemm 256x4096x4096 on 16 threads ran 1.13 to 1.20 times Intel MKL. On the other sizes we measured it ran slower than MKL, so no parity is claimed. | Partial | 1.5.0 | Householder 1958; Bowdler et al. 1968; Cuppen 1981; Gu, Eisenstat 1995 |
| GPU J and K buildsDirect screened Coulomb and exchange on the GPU up to d functions, and GPU RI-J three-centre integrals, in CUDA builds. Off by default until measured faster than the CPU. | Unverified | 1.5.0 | Ufimtsev, Martinez 2008; Ufimtsev, Martinez 2009; Haser, Ahlrichs 1989 |
Printed Value, Our Value
| Quantity | Source | Printed | Ours | Result |
|---|---|---|---|---|
| Silicon a (A), c11, c12, c44 Clamped, c44 Relaxed, B (Mbar); Different Pseudopotential | Nielsen, Martin 1985 Table I, doi 10.1103/PhysRevB.32.3792 | 5.400, 1.59, 0.61, 1.11, 0.85, 0.93 | 5.379, 1.638, 0.644, 1.101, 0.804, 1.003 | Pattern |
Exact Checks
Our output against exact mathematics or a recomputation inside the engine, never another program. These are not printed-value checks and are counted separately.
| Quantity | Checked Against | Tolerance | Ours | Result |
|---|---|---|---|---|
| Distributed SCF Energy on 2 and 3 Ranks: Water RHF, Methyl Radical UHF and Ethane RHF in 6-31G | The same SCF on one rank | 1e-10 Eh | worst 1.1e-13 Eh; the same rank count twice is identical bit for bit | Within tolerance |
| Water HF/STO-3G Frequencies and Infrared Intensities from Nine Hessian Rows Computed in Three Parts and Merged | One run with the same Cartesian displacements | 0 (bit for bit) | identical | Within tolerance |
Not Offered Today
- Hosted compute (runs only on servers you add)
- Real-network speed-up claims until measured
- Windows HPC head nodes
- A Slurm cluster provisioned by AiMolSim on your Linux servers (design stage, nothing built)
- MPI between servers (the engine reports MPI off)
In the Chemistry Lab
Cluster, then Add cluster: pick your own servers, confirm its key fingerprint, then turn on Run on cluster in any computation. The cluster page shows the queue, usage by person, history with a CSV download and a printable monthly report; servers can carry labels and maintenance windows.
HPC and Clusters Questions
Do you run jobs on your own cloud?
How are cluster jobs ordered?
Is a distributed result the same as a single-server result?
See It on a Molecule of Your Own
Start with one molecule and one energy. Every number comes with its method, its basis, its engine version and the server it ran on.