Skip to content

The accuracy-time frontier

GPU accuracy-vs-time frontier at 16M

Accuracy vs fit time at 16M rows

variant iters fit_s test r2
bonsai_cuda_depthwise 60 7.79 0.8459
bonsai_cuda_depthwise 80 9.81 0.8669
bonsai_cuda_depthwise 100 11.88 0.8777
bonsai_cuda_depthwise 130 14.84 0.8859
bonsai_cuda_depthwise 200 21.08 0.8918
bonsai_cuda_depthwise 300 29.21 0.8934
bonsai_cuda_oblivious 60 7.11 0.8386
bonsai_cuda_oblivious 80 8.77 0.8613
bonsai_cuda_oblivious 100 10.34 0.8743
bonsai_cuda_oblivious 130 12.65 0.8852
bonsai_cuda_oblivious 200 17.50 0.8946
bonsai_cuda_oblivious 300 24.09 0.8973
bonsai_cuda_oblivious 450 33.10 0.8979
xgb_cuda 100 20.25 0.8779
xgb_cuda 150 23.55 0.8887
xgb_cuda 200 26.16 0.8918
xgb_cuda 300 31.98 0.8933
catboost_gpu 100 18.54 0.8751
catboost_gpu 150 22.40 0.8892
catboost_gpu 200 26.32 0.8944
catboost_gpu 300 34.00 0.8973
catboost_gpu 450 45.09 0.8980

Source: gpu-pareto-16M-2026-07.jsonl. bonsai is first to every measured accuracy across the grid (terminal accuracies tie within the noise band) and its marginal round cost stays below CatBoost's on the same pod. Evidence: benchmarks/gpu-pareto-16M-2026-07.md. Measured at d3ffcd0 (2026-07-31, pod-NVIDIA-L40S).

Ordered boosting at scale (CatBoost door)

The probe behind decisions 62 to 64: CatBoost's Ordered vs Plain modes against bonsai oblivious as rows grow.

door rows learner knob fit_s test r2
ordered 200,000 catboost_ordered iters=100 17.06 0.8722
ordered 200,000 catboost_plain iters=100 2.51 0.8720
ordered 200,000 bonsai_oblivious iters=100 3.91 0.8747
ordered 200,000 catboost_ordered iters=200 30.59 0.8932
ordered 200,000 catboost_plain iters=200 4.66 0.8937
ordered 200,000 bonsai_oblivious iters=200 7.30 0.8940
ordered 1,000,000 catboost_ordered iters=100 63.22 0.8754
ordered 1,000,000 catboost_plain iters=100 8.36 0.8760
ordered 1,000,000 bonsai_oblivious iters=100 9.98 0.8766
bins 1,000,000 catboost_plain bin samples=None 8.01 0.8760
bins 1,000,000 bonsai_oblivious bin samples=200000 9.76 0.8766
bins 1,000,000 bonsai_oblivious bin samples=1000000 9.98 0.8766
bins 4,000,000 catboost_plain bin samples=None 36.73 0.8762
bins 4,000,000 bonsai_oblivious bin samples=200000 34.36 0.8763
bins 4,000,000 bonsai_oblivious bin samples=1000000 35.40 0.8766
bins 4,000,000 bonsai_oblivious bin samples=4000000 36.62 0.8764
isolate 16,000,000 catboost_plain_cpu - 157.67 0.8744
isolate 16,000,000 bonsai_oblivious_cpu - 171.75 0.8749
isolate 16,000,000 bonsai_depthwise_cpu - 151.93 0.8782
gpu 16,000,000 bonsai_cuda_oblivious_prefix iters=100 - 0.8638
gpu 16,000,000 bonsai_cuda_oblivious_fixed iters=100 29.21 0.8749
gpu 16,000,000 bonsai_cuda_oblivious_fixed iters=160 41.03 0.8913
gpu 16,000,000 bonsai_cuda_depthwise iters=100 30.37 0.8776
gpu 16,000,000 catboost_gpu iters=100 23.81 0.8751
gpu 16,000,000 catboost_gpu iters=150 28.37 0.8892
gpu 16,000,000 catboost_gpu iters=200 32.19 0.8944

Source: catboost-scale-edge-2026-07.jsonl. Evidence: benchmarks/catboost-scale-edge-2026-07.md.