ashvardanian/ScalingElections: v0.2.0

Le résumé fourni par la source

Tiled parallel implementations of the Schulze voting method across CPUs and GPUs, written in Mojo and in CUDA C++ wrapped into Python. The method is used by Pirate Parties and open-source foundations, and its Floyd-Warshall-shaped inner loop is a good example of a combinatorial problem parallelized by reordering evaluation rather than by changing the result. Backends CUDA C++ for NVIDIA, with Hopper TMA tiles, targeting the local device's compute capability ROCm/HIP for AMD, sharing the same sources Mojo, a single file covering CPU, SIMD and GPU paths OpenMP host kernels with #pragma omp simd tiles and explicit Arm NEON intrinsics Numba and NumPy serial references for validation No CMake: the native build is packed into setup.py, and pixi drives the Mojo side. Throughput, in pairwise candidate comparisons per second | Candidates | Numba 384c | Mojo 384c | Mojo SIMD 384c | CUDA h100 | Mojo h100 | Mojo mi355x | | :--------- | -----------: | ----------: | ---------------: | ----------: | ----------: | ------------: | | 8'192 | 74.6 gcs | 76.6 gcs | 357.3 gcs | 495.3 gcs | 408.0 gcs | 2.4 tcs | | 16'384 | 76.7 gcs | 80.7 gcs | 369.0 gcs | 600.7 gcs | 635.3 gcs | 2.9 tcs | | 32'768 | 101.4 gcs | 82.3 gcs | 293.1 gcs | 921.4 gcs | 893.7 gcs | 3.5 tcs | Measured on dual-socket Xeon 6 with 384 cores, NVIDIA H100, and AMD MI355X. Background and derivation: https://ashvardanian.com/posts/scaling-elections/

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.