Largest tables the browser handles comfortably; above these, fit throws TOO_LARGE.
Cost grows with training rows x features: measured on onnxruntime-node CPU with 8
members and 100 test rows, 500x10 took 2.5 s, 1000x10 5.4 s, 2000x20 67 s, 1000x50 91 s.
WebGPU is faster; fewer ensemble members (nEstimators) cut time proportionally.
Largest tables the browser handles comfortably; above these, fit throws TOO_LARGE. Cost grows with training rows x features: measured on onnxruntime-node CPU with 8 members and 100 test rows, 500x10 took 2.5 s, 1000x10 5.4 s, 2000x20 67 s, 1000x50 91 s. WebGPU is faster; fewer ensemble members (nEstimators) cut time proportionally.