Benchmarks
Performance comparisons of Aicraft against other frameworks.
Forward Pass Latency (MNIST, batch=1)
| Framework | Time | Binary Size |
|---|---|---|
| Aicraft (AVX2) | 0.42 ms | ~150 KB |
| Aicraft (scalar) | 1.8 ms | ~120 KB |
| PyTorch | 2.1 ms | ~800 MB |
| TensorFlow Lite | 1.5 ms | ~5 MB |
Training Throughput (MNIST, 60k samples)
| Framework | Epoch Time | Memory |
|---|---|---|
| Aicraft (AVX2) | 3.2 s | 8 MB |
| Aicraft (Vulkan) | 1.8 s | 12 MB |
| PyTorch | 4.5 s | 450 MB |
| TensorFlow | 5.1 s | 1.2 GB |
Comparison Summary
| Metric | Aicraft | PyTorch | TensorFlow |
|---|---|---|---|
| Binary size | ~150 KB | ~800 MB | ~1.8 GB |
| Dependencies | 0 | ~50 | ~80 |
| Language | C11 | C++ / Py | C++ / Py |
| GPU backend | Vulkan | CUDA | CUDA |
| SIMD | Hand-tuned | Generic | Generic |
| Memory | Arena allocator | malloc/free | Custom |
| Edge deploy | MCU-ready | No | TFLite |
Test Environment
- CPU: Intel Core i7-12700K
- GPU: NVIDIA RTX 3060 (Vulkan)
- RAM: 32 GB DDR5
- OS: Ubuntu 22.04 LTS
- Compiler: GCC 12.3 with
-O3
note
Benchmarks are indicative. Your mileage may vary depending on hardware and workload.