View this project on GitHub ↗
MNIST end-to-end on an A100
Architecture + optimizer + kernel codesign
- Learn to read digits from 10,000 labelled 9x9 MNIST images, then label 10,000 more, from scratch, on one A100.
- Every call runs in a fresh, sandboxed process on a fresh draw, whitened and secretly rotated.
- Ranked by time per call. Energy per call above idle is reported beside it (how).

API
import mnist
def my_method(train_x, train_y, test_x):
# train_x (10000, 60) float32, train_y (10000,) int64 in 0..9, test_x (10000, 60) float32, on the GPU
...
return labels # (10000,) integer tensor on the GPU
if __name__ == "__main__":
ms = mnist.score(my_method, difficulty=1) # 1 (loosest) to 5; ms per call, or raises mnist.Disqualified
python example.py # on your own GPU
python run_modal.py example.py:mlp --difficulty 1 --runs 3 # official: three Modal A100-80GB runs, median
The rules are in mnist.py’s docstring.
Each row is a separate submission by the listed entrant; source attribution and
verification scope are in its report. Rows with one verification run are
provisional. ★ identifies an established record. Non-record submissions are welcome.
Research with failed or incomplete qualification checks can be included with all
outcomes documented in the research submissions below;
its inclusion does not qualify it for a difficulty table.
Difficulty = 1 (error ≤ 5.40%)
| Date |
mJ |
ms |
Submission |
Entrant |
Description |
| 2026-09-25 |
29,662 |
700.1 |
py, report |
@yaroslavvb |
example.py baseline: MLP 60-1024-1024-10, 400 steps |
| 2026-09-26 |
2,648 |
191.3 |
py, report |
@yaroslavvb |
MLP 60-1024-1024-10, 100 steps |
| 2026-09-26 |
2,131 |
61.7 |
py, report |
@yaroslavvb |
MLP 60-256-256-10, 200 steps in a CUDA graph |
| 2026-09-30 |
3,634 |
351.0 |
py, report |
@yaroslavvb |
MLP, 200 steps, TF32 and fused AdamW; one sandboxed run |
| 2026-10-08 |
530 |
6.7 |
py, report |
@sjbaebae |
Two 60-256-256-10 MLPs, 30 updates, three-neighbor voting; median of three sandboxed Modal runs |
| 2026-10-12 |
60 |
1.6 |
py, report |
@spartanhaden |
Nyström RBF kernel ridge in a Fisher metric, 480 class-wise k-means landmarks, one Cholesky solve; no backprop; median of three independent reviewer sandboxed Modal runs ★ best |
Difficulty = 2 (error ≤ 3.40%)
| Date |
mJ |
ms |
Submission |
Entrant |
Description |
| 2026-09-26 |
200,870 |
927.3 |
py, report |
@yaroslavvb |
16 MLPs 60-1024-1024-10, 400 steps |
| 2026-09-26 |
14,436 |
252.1 |
py, report |
@yaroslavvb |
4 MLPs 60-256-256-10, 800 steps in a CUDA graph |
| 2026-09-29 |
8,071 |
38.3 |
py, report |
@islamborghini |
Non-neural RBF kernel ridge, one learned metric update, 16-step Nyström-preconditioned CG |
| 2026-09-30 |
132,635 |
761.9 |
py, report |
@yaroslavvb |
8 MLPs, 500 steps, EMA 0.992; one sandboxed run |
| 2026-10-08 |
1,333 |
14.6 |
py, report |
@sjbaebae |
Two 60-256-256-10 MLPs, 120 updates, three-neighbor voting; median of three sandboxed Modal runs |
| 2026-10-12 |
603 |
4.6 |
py, report |
@spartanhaden |
RBF kernel ridge with one learned (AGOP) metric, 5 Nyström-preconditioned CG steps; no backprop; median of three independent reviewer sandboxed Modal runs |
| 2026-10-19 |
208 |
3.2 |
py, report |
@spartanhaden |
Fisher-metric RBF ridge on a compressed weighted training set, triangular FITC-style preconditioning and extrapolated PCG; ports sigkillme0’s D3 method; median of three independent reviewer sandboxed Modal runs |
| 2026-10-22 |
172 |
2.9 |
py, report |
@spartanhaden |
Fisher-PCG D2 with a standard-form Woodbury preconditioner, FP16 products, side-stream class moments and batched small factors; learner constants unchanged; median of three independent reviewer sandboxed Modal runs ★ best |
Difficulty = 3 (error ≤ 2.70%)
| Date |
mJ |
ms |
Submission |
Entrant |
Description |
| 2026-09-28 |
1,159,909 |
15,818.8 |
py, report |
@SethTS |
Ladder 60-1000-500-250-250-250-10, 5,000 steps in a CUDA graph |
| 2026-09-28 |
676,850 |
6,556.9 |
py, report |
@SethTS |
Ladder, 6,500 steps, fused Triton kernels in a CUDA graph |
| 2026-09-29 |
259,847 |
1,895.4 |
py, report |
@SethTS |
Ladder, batch 1,000, 1,200 steps, TF32, fused Triton kernels in a CUDA graph |
| 2026-09-29 |
201,792 |
1,443.3 |
py, report |
@SethTS |
Ladder, batch 1,000, 900 steps, labelled rows sampled by loss, TF32, fused Triton kernels in a CUDA graph |
| 2026-09-30 |
253,761 |
1,732.2 |
py, report |
@yaroslavvb |
Ladder, 1,100 steps; 8.3% fewer updates; one sandboxed run |
| 2026-10-05 |
151,404 |
1,157.1 |
py, report |
@islamborghini |
Sampled Ladder, batch 1,000, 800 steps, TF32; tuned forward and backward Triton kernels; median of three sandboxed runs |
| 2026-10-08 |
98,226 |
416.1 |
py, report |
@sjbaebae |
Eight 60-256-256-256-10 RMSNorm/ReLU MLPs, 1,250 updates, five-neighbor voting and three-step query graph propagation; median of three sandboxed Modal runs |
| 2026-10-12 |
8,839 |
51.8 |
py, report |
@sjbaebae |
Spherical direction/radius RBF, two tangent-gradient metric updates; derived from PR #116; median of three independent reviewer sandboxed Modal runs |
| 2026-10-12 |
2,362 |
18.1 |
py, report |
@spartanhaden |
RBF kernel ridge, two AGOP metric updates, HKNN local-hull correction on the 512 lowest-margin rows; no backprop; median of three independent reviewer sandboxed Modal runs |
| 2026-10-14 |
7,899 |
44.0 |
py, report |
@islamborghini |
RBF kernel ridge, two analytic metric updates, CUDA-graph Nyström-preconditioned CG and local class geometry; median of three independent reviewer sandboxed Modal runs |
| 2026-10-18 |
320 |
5.2 |
py, report |
@sigkillme0 |
Learned-metric RBF kernel with fused Triton PCG, compressed kernel tiles, and local-hull correction; builds on Haden’s RFM/HKNN learner; median of three independent reviewer sandboxed Modal runs |
| 2026-10-19 |
639 |
4.6 |
py, report |
@sjbaebae |
Fisher-PCG with 6,144 hard rows, eight PCG steps and native FP32 local-hull CG; builds on Haden, sigkillme0 and Islam’s kernel work; median of three independent reviewer sandboxed Modal runs |
| 2026-10-22 |
359 |
3.5 |
py, report |
@spartanhaden |
Plain-PyTorch Fisher-PCG with a rank-48 metric, natural-gradient step, Woodbury preconditioner and HKNN on compacted low-margin rows; median of three independent reviewer sandboxed Modal runs ★ best |
Difficulty = 4 (error ≤ 2.30%)
| Date |
mJ |
ms |
Submission |
Entrant |
Description |
| 2026-09-28 |
1,143,255 |
11,094.6 |
py, report |
@SethTS |
Ladder, 11,000 steps, fused Triton kernels in a CUDA graph |
| 2026-09-29 |
547,382 |
3,784.2 |
py, report |
@SethTS |
Ladder, batch 1,000, 2,400 steps, TF32, fused Triton kernels in a CUDA graph |
| 2026-09-29 |
400,682 |
2,882.5 |
py, report |
@SethTS |
Ladder, batch 1,000, 1,800 steps, labelled rows sampled by loss, TF32, fused Triton kernels in a CUDA graph |
| 2026-09-30 |
460,763 |
3,481.3 |
py, report |
@yaroslavvb |
Ladder, 2,200 steps; 8.3% fewer updates; one sandboxed run |
| 2026-10-12 |
63,113 |
580.1 |
py, report |
@spartanhaden |
Ladder MLP 60-1024-512-256-10 on AGOP-metric inputs, 560 Adam steps, kernel-ridge teacher; median of three independent reviewer sandboxed Modal runs |
| 2026-10-18 |
42,617 |
219.2 |
py, report |
@islamborghini |
Five RBF classifiers with three analytic AGOP metric updates each, Nyström-preconditioned CG, transductive refits and HKNN affine-hull correction; no backprop; median of three independent reviewer sandboxed Modal runs |
| 2026-10-19 |
14,754 |
97.0 |
py, report |
@spartanhaden |
Two RFM teachers distilled into three tangent-jittered MLP students with cross-teaching and HKNN; one full-call CUDA graph; median of three independent reviewer sandboxed Modal runs |
| 2026-10-19 |
12,845 |
78.3 |
py, report |
@sjbaebae |
Haden’s tangent hybrid with 250 Adam updates and cross-teaching at 80/170 steps; teacher, students and readout unchanged; median of three independent reviewer sandboxed Modal runs |
| 2026-10-22 |
6,664 |
38.7 |
py, report |
@spartanhaden |
Two Fisher-PCG teachers distilled into three tangent-jittered MLP students; 220 updates with a fused Triton student step and HKNN; median of three independent reviewer sandboxed Modal runs |
| 2026-10-25 |
3,347 |
25.3 |
py, report |
@sjbaebae |
Fisher-PCG teacher distilled into three width-128 tangent students, split-four gradients and rolled Triton row loops; retains Haden and upstream kernel credits; median of three independent reviewer sandboxed Modal runs |
| 2026-11-01 |
2,259 |
22.1 |
py, report |
@spartanhaden |
Fisher-PCG teacher with three width-128 tangent students, class-wise neighbors, without-replacement row walks, 1,280 labelled and 256 unlabelled rows, and 220 fused Triton updates; retains #140 and upstream credits; median of three independent reviewer sandboxed Modal runs ★ best |
Difficulty = 5 (error ≤ 1.90%)
| Date |
mJ |
ms |
Submission |
Entrant |
Description |
| 2026-09-28 |
4,410,730 |
36,201.2 |
py, report |
@SethTS |
Ladder, 36,000 steps, fused Triton kernels in a CUDA graph |
| 2026-09-29 |
2,059,576 |
14,130.6 |
py, report |
@SethTS |
Ladder, batch 1,000, 9,000 steps, TF32, fused Triton kernels in a CUDA graph |
| 2026-09-30 |
1,824,760 |
13,264.3 |
py, report |
@yaroslavvb |
Ladder, 8,400 steps; 6.7% fewer updates; one sandboxed run |
| 2026-10-12 |
638,234 |
3,335.4 |
py, report |
@likelytobelaura |
Eight RMSNorm/GELU MLPs, Muon/AdamW, 2,500 updates, local tangent jitter, disjoint-pool cross-teaching and graph readout; median of three independent reviewer sandboxed Modal runs |
| 2026-10-12 |
226,352 |
1,818.1 |
py, report |
@spartanhaden |
Ladder MLP 60-1024-512-256-10 on AGOP-metric inputs, 2,300 Adam steps, kernel-ridge teacher; median of three independent reviewer sandboxed Modal runs |
| 2026-10-14 |
343,152 |
1,915.4 |
py, report |
@likelytobelaura |
Eight width-256 RMSNorm/ReLU MLPs, Muon, tangent jitter, cross-teaching and graph readout; median of three independent reviewer sandboxed Modal runs |
| 2026-10-18 |
154,289 |
897.5 |
py, report |
@islamborghini |
Haden’s kernel-taught Ladder tuned to batch 2,000 and 800 steps; architecture, teacher and optimizer unchanged; median of three independent reviewer sandboxed Modal runs |
| 2026-10-19 |
107,886 |
633.5 |
py, report |
@spartanhaden |
Kernel-taught Ladder with tangent jitter, hard-query sampling and fused Triton combinator backward; 600 Adam steps of batch 2,500; median of three independent reviewer sandboxed Modal runs |
| 2026-10-19 |
77,090 |
492.8 |
py, report |
@sjbaebae |
Haden’s kernel-taught Ladder narrowed to 768-512-256, 480 native fused Adam updates and batch 2,500; teacher and branchless guard retained; median of three independent reviewer sandboxed Modal runs |
| 2026-10-22 |
67,813 |
406.9 |
py, report |
@spartanhaden |
Kernel-taught width-768 Ladder, 450 Adam updates, labelled-row top-down targets, per-call noise bank and hidden-space HKNN; median of three independent reviewer sandboxed Modal runs |
| 2026-10-25 |
40,390 |
273.9 |
py, report |
@sjbaebae |
Width-480 label-top-down Ladder, 315 updates, per-call transpose/near-identity initialization, population prediction statistics, parallel gradient reductions and decoder/hull readout; retains Ladder and kernel lineage credits; median of three independent reviewer sandboxed Modal runs |
| 2026-11-01 |
30,157 |
226.9 |
py, report |
@spartanhaden |
Width-480 Ladder with decoupled weight decay on eight encoder/decoder matrices and 240 AdamW updates; retains #140 and upstream Ladder/kernel credits; median of three independent reviewer sandboxed Modal runs ★ best |
Research submissions
These non-record experiments are included with their original runtime code and
author evidence. The outcomes below come from the retained independent reviewer
runs on those frozen sources. Each report separates author results from reviewer
results and preserves failed and incomplete runs. Research inclusion does not
qualify a result for a difficulty table or establish a record.