RunMat: GPU Computing Platform for Engineering Math
GPU-accelerated MATLAB-syntax math with a built-in agent that runs code, checks results, and iterates. Open source, browser or CLI. No license required.
Run math blazing fast
MATLAB-syntax math on GPU, in your browser or from the CLI. With the built-in agent, sweep fifty parameter variations in the time it used to take for one.
GPU accelerated plotting
Render millions of points straight from a GPU tensor. 2D and 3D plots run on the same compute chain as your math, in the same runtime. The platform's agent has its own camera and can rotate, zoom, and inspect plots while you work on something else.
An engineering team, working alongside you
The agent runs your scenarios, reviews the code, and searches the project for what you need. You explore problems that used to take a team.
Runtime-aware inspection
Checks tensors for NaN, verifies shapes against expected dimensions, and reads plot data from any 3D camera angle.
Every change is reviewable
Edits are presented as diffs. Accept or reject each change. Conversations are stored as searchable project files.
Syntax you already know.
MATLAB syntax reads like the whiteboard: one line of math, one line of code. You and your team already know it. Write the math you mean, and RunMat routes it across CPU and GPU automatically.
The fastest runtime for your math
In verified benchmarks, RunMat runs 150-180x faster than GNU Octave, up to 10x faster than NumPy, and up to 3.8x faster than PyTorch on GPU image pipelines. The runtime fuses sequential operations into fewer GPU steps and keeps your arrays on-device between them.
Debug with full visibility
Debug faster by seeing everything as you write. Hover any variable to see its shape and type. Click on an intermediate value to inspect it. Dimension mismatches are flagged in the editor before you run. The agent uses the same inspection tools: materializing variables, checking shapes, and locating where a computation went wrong.
Every change versioned. No git required.
Every save is a version, automatically. Per-file history and full project snapshots track every change, even on terabyte-scale datasets. Share projects with your team, no git setup, no merge conflicts. Agent-generated edits get the same treatment: versioned and reversible.
Automatic file history
Every save creates a version. Browse the timeline, restore any previous state. No commits, no staging.
Project snapshots
Capture your entire project in one click. Restore instantly. A clean timeline with no merge conflicts.
App project sharing
Share projects with colleagues instantly. No shared drives, no emailing files.
Scale your math, not your toolchain
Same code runs on Apple, Nvidia, and ARM GPUs across macOS, Windows, and Linux. For large data, a sharded cloud filesystem handles multi-petabyte datasets with parallel reads and writes designed for NIC saturation. Delta snapshots version your datasets efficiently without duplicating terabytes.
Any GPU
Metal on Mac, Vulkan on Linux and ARM, DirectX 12 on Windows. No CUDA dependency.
Any OS
macOS, Windows, Linux, and headless servers. Same runtime, same results.
High-bandwidth cloud filesystem
Sharded, petabyte-scale storage with parallel I/O and delta snapshots for efficient versioning.
Open source, MIT licensed
Read every line of code that runs your math. Fork the runtime to audit or self-host. No vendor lock-in, no black boxes. The runtime is on GitHub and actively maintained. See how we validate every numerical path.
Secure by design
Security, compliance, and deployment options for teams that protect proprietary research and engineering data.
SSO & SCIM
Integrate with your identity provider. Provision and deprovision users automatically.
ITAR-compliant deployment
Self-hosted, air-gapped option available for export-controlled environments.
Open source & auditable
MIT-licensed runtime. Inspect every line of code that runs your math.
SOC 2 ready
Built to SOC 2 standards. Audit planned for Q4 2026.
Sandboxed agent execution
Agent commands run inside OS-level sandboxes (Seatbelt on macOS, bubblewrap on Linux), with network denied by default, filesystem scoped to your project, and known secret patterns stripped.
Replayable agent sessions
Durable sessions journal each agent step, so teams can replay work for audit and review.
Provider-agnostic architecture
The agent runs on a model-agnostic harness. No lock-in to one provider's roadmap.
Try RunMat for free
Write code or describe what you want to compute. The sandbox is free, no account required.