// Coifman–Lafon · sparse eigensolver · B300 GPU
Diffusion maps at 100M-sample scale. The intrinsic geometry of your data — conformational landscapes, reaction coordinates, latent structure — in hours, not weeks.
The DAC paper showed that FlashAttention and diffusion maps are two regimes of the same Markov geometry. FlashDiffusion is the equilibrium brother.
Asymmetric kernel — query and key differ. Non-equilibrium steady state. Powers transformers and language models.
Symmetric kernel — query equals key. Equilibrium. Recovers the intrinsic geometry of scientific data.
Eigenvector compute time as dataset size grows. FlashDiffusion vs. dense O(N²D) without the flash kernel.
Geometric analysis first. Guided learning second.
MD trajectories (.xtc), cryo-EM stacks, XFEL diffraction, or any high-dimensional scientific data. Stored on R2.
R2 · no size limitSparse Coifman–Lafon diffusion map. B300-native flash kernel computes the intrinsic geometry of your data.
B300 · hours not weeksTrain a neural network guided by the eigenvectors — MLP, DiT, or autoregressive. Physically meaningful, not a black box.
MLP · DiT · ARConformational landscapes of proteins, polymers, and materials. Identify slow collective variables from millions of frames.
Heterogeneity analysis of particle stacks. Map conformational continua without discrete class assumptions.
Single-particle and serial crystallography datasets. Geometry-aware analysis of femtosecond snapshots.
Private beta.
Apply now.
If you have a dataset and a problem that needs diffusion maps at scale, we want to hear from you.
// No spam. We respond personally.