"Optimizing Scientific Python with C++, CUDA, and Serverless Compute"
summary
Migration story with numbers and tools: moving algorithms that touch tens of millions of points into serverless where Lambda has ~10 GB RAM and short timeouts forces rethinking, not lift‑and‑shift. Profiling workflow uses PySpy, Intel VTune, NVIDIA Nsight and Valgrind; C++ bindings via pybind11 and Eigen for SIMD and Intel ICX with SVML; CUDA implements batched LU/Cholesky on GPU; claims 3–10× measured speedups and CI guards that validate every fast path against a trusted reference.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:12Desktop to cloud challenges
- 3:19Python with C++ bottlenecks
- 5:26Profiling tools
- 7:30Memory allocation optimization
- 9:19Python-C++ integration
- 12:26Compiler tuning
- 15:38Floating-point precision
- 16:21SIMD vectorization
- 18:46Vectorization libraries
- 19:26GPU acceleration with CUDA
- 20:51Cloud benefits and limits
- 21:57RAM in serverless
- 22:39Cloud traceability
- 23:42Correctness checks
- 25:28CI for C++/CUDA
- 26:11Optimization strategies
- 26:53Summary
- 27:25Q&A
related talks
30mFrom Matlab to Lambda: Transforming Structural Engineering Research Tools
25m"Boosting Python Numeric Computations with Dynamic C++ Integration"
30mCUDA in Your Python: Effective Parallel Programming on the GPU
30mHighLoad Python SIMD, GPU, and Horizontal Scaling Looking into Silicon
36mAn almost perfect match: cloud native and scientific computing
02h 28mGPU Accelerated Python
46mServerless for Data Scientists
27mSpeed up Python with hardware accelerators