CIFAR-10 CUDA models
CUDA reproduction of the CIFAR-10 architecture matrix, trained on an NVIDIA A40 with batch size 1024, seed 42, GPU-resident tensor shards, GPU-side augmentation, and a 200-epoch target (R-CNN-B uses 30 epochs). JSON files are the authoritative per-run metadata; checkpoints are provided alongside them.
Selected results
| Model | Best / target | Best validation | Test accuracy | Parameters |
|---|---|---|---|---|
| ResNet-50 | 191/200 | 92.60% | 91.80% | 23,520,842 |
| DenseCNN | 191/200 | 92.10% | 91.50% | 300,634 |
| CNN-C + BN | 200/200 | 89.00% | 88.35% | 2,360,906 |
| ViT | 174/200 | 81.24% | 80.15% | 546,186 |
The complete matrix is documented in the GitHub report. All training curves and run configuration are public in the W&B CUDA project. The website includes a CUDA comparison section.
These results are intentionally separate from the MPS reference leaderboard: the accelerator, batch size, and input pipeline differ, so a CUDA/MPS gap is not evidence of a hardware-only effect.
Attention analysis
The CUDA DenseCNN checkpoint was inspected with the same Grad-CAM procedure as
the MPS analysis. The target layer is DenseCNN.features.5; the channel panels
rank the top 10 channels by absolute gradient-weighted mean activation.
| Test index | True label | DenseCNN prediction | CNN-C + BN prediction | Interpretation |
|---|---|---|---|---|
| 0 | cat | cat (97.2%) | cat (100.0%) | Both models focus on a correctly classified object. |
| 1 | ship | ship (95.9%) | ship (100.0%) | Both correctly identify the dominant ship evidence. |
| 3957 | truck | airplane (89.8%) | airplane (85.2%) | Shared false positive: background/shape evidence dominates class evidence. |
| 5808 | dog | cat (95.0%) | horse (60.4%) | Both fail on a fine-grained animal; the low-resolution crop is ambiguous. |
Comparison heatmap: DenseCNN vs CNN-C + BN
Top-channel panels: index 0 · index 1 · index 3957 · index 5808
| Correct cat | Correct ship |
|---|---|
 |
 |
| Shared truck → airplane false positive | Dog failure case |
|---|---|
 |
 |
The current artifact was rendered with CPU visualization because the publishing environment has neither CUDA nor MPS available; the weights and predictions are from the CUDA-trained checkpoints. This affects rendering speed, not the model weights or selected class.