// nuwright
◔ building · groundwork done; no stage built yet

MoMap

GPU-accelerated UMAP for Apple Silicon, written in Mojo and compatible with umap-learn.

job
GPU UMAP embedding
domain
developer tools · dimensionality reduction
runs
local
class
library
since
2026

One call

  1. 01kNNNearest neighbours on the GPU
  2. 02graphA sparse, symmetric neighbour graph
  3. 03initA starting layout
  4. 04layoutGPU optimisation with exact atomic updates
  5. 05outputCoordinates, or a map file

In · out

inputs
  • a table of numbers (a NumPy array or .npy file)
outputs
  • 2D or 3D coordinates (planned)
  • a map file other tools can open (planned)

What is measured

Against: umap-learn 0.5.12 baseline on MNIST, Fashion-MNIST, COIL-20

Target: at least 5× the umap-learn baseline on MNIST, on an M3 Ultratarget
The bar to beat: umap-learn takes 4.5 s on MNISTmeasured (baseline, not MoMap)
GPU groundwork: unsynchronised updates lose 76 to 92% of writes, while atomic adds are exact, so the layout uses atomicsmeasured

Stack · links

MojoPythonumap-learn (reference)scikit-learn
last updated 2026-09-28 · from source, not marketingbuilding · no result is claimed