First Interstellar Institute
The physics of reaching the stars.
We test warp drives, traversable wormholes and negative-mass propulsion against Einstein’s equations — evolved in full numerical relativity on GPUs, and published openly with the code behind every result.
01The Institute
No propulsion we know how to build can carry people to another star within a human lifetime. The First Interstellar Institute asks whether physics allows one that could — and answers with the same numerical-relativity machinery used to model colliding black holes.
Negative mass
Bondi’s self-accelerating pair of positive and negative mass, evolved in full 3+1 numerical relativity.
Read the paperTraversable wormholes
Ellis–Bronnikov wormholes alone and in pairs: collapse, inflation, mergers and the gravitational waves they emit.
Latest paperWarp metrics
Alcubierre-class spacetimes, the energy they demand, and open tooling to analyse them.
Warp drive lectureGPU numerical relativity
Einstein’s equations solved on GPUs with GRTeclyn, checked by convergence tests and null runs.
How we work- Papers, each with its code and data
- 3
- Full numerical-relativity evolutions behind the binary-wormhole paper
- 103
- NVIDIA H100 GPU-hours to run them
- 592
- Free lectures, from Einstein’s equations to warp metrics
- 4
02The problem
At the speeds we know, the stars are out of reach.
The distances between stars are so vast that even our most advanced propulsion concepts require centuries. Only a breakthrough at the frontier of physics can make interstellar travel practical within a human lifetime.
- Nuclear electric — known physics, known engineering
- Fusion — known physics, unknown engineering
- Warp — unknown physics, unknown engineering
03How we work
AI made GPUs cheap. GRTeclyn makes them solve Einstein’s equations.
Testing a warp drive or a wormhole means solving Einstein’s equations in full, a job that used to need time on a national supercomputer. Two things changed.
Up to
20× faster
The same simulation, the same code, on GPUs instead of CPUs.
Our 103-run binary-wormhole campaign
- On CPUs
- up to ~16 months
- On one NVIDIA H100
- ~25 days
CPU cluster. Each node keeps its own memory, so data crawls between nodes and stalls (red).
One GPU. Thousands of cores share one fast memory pool, so everything moves at once (blue).
| Why the GPU wins | CPUServer CPU, 64-core | H100NVIDIA H100 | Gain |
|---|---|---|---|
| Parallel cores | ~64 | ~16,900 | ~260× |
| FP64 throughput | ~4 TFLOP/s | ~34 TFLOP/s | ~8× |
| Memory bandwidth | ~0.5 TB/s | ~3.35 TB/s | ~7× |
Every timestep, each grid cell reads its neighbours, so the update is limited by memory bandwidth. That ~7× is the number that matters most.
Riding the AI wave
The chips that train language models, H100s with HBM3 memory, are exactly what a numerical-relativity code needs, and AI demand has turned them into something you rent by the hour. Our 103-run binary-wormhole campaign took 592 H100 GPU-hours, with no supercomputer allocation.
GRTeclyn, adapted
GRTeclyn is the GRTL Collaboration’s open-source successor to GRChombo, rebuilt on AMReX to run on GPUs, up to 20× faster than on CPUs. It evolves Einstein’s equations in the CCZ4 formulation with adaptive mesh refinement. We extended it with exotic matter, phantom scalar fields for wormholes and negative-mass stars for the Bondi dipole, and publish the code.
“A planet is the cradle of the mind, but one cannot live in a cradle forever.”
