Scale Smarter, Not Harder, with cuPyNumeric.
summary
cuPyNumeric claims to be a drop-in replacement for NumPy that automatically shards work across CPUs and GPUs using the Stanford task-based runtime (Legate). The uploader reports a real-world SLAC workload moving from single-node NumPy to cuPyNumeric and achieving 6× speed-up on multi-node GPUs; the transcript adds faster examples like full detector analysis in 5 minutes versus 10+ hours. Code changes are mostly renames but require nontrivial tuning; the library integrates with SciPy, matplotlib and JAX.
AI-written summary · about
chapters (8)
- 0:00Introduction to cuPyNumeric
- 1:05Material science needs GPU
- 7:12cuPyNumeric accelerates X-ray data
- 12:12X-ray free electron laser experiment
- 16:27Peta-scale data analysis challenge
- 19:00Live experiment feedback with cuPyNumeric
- 21:31Future impact on discovery
- 22:34Q&A
related talks
30mCuPy: A NumPy-compatible Library for GPU
39mAccelerated Data Science Libraries
20mNumS: Scalable Array Programming for the Cloud
02h 28mGPU Accelerated Python
02h 55mGPU Accelerated Python w: CUDA
27mSpeed up Python with hardware accelerators
35mCUDA in Python- A New Era for GPU Acceleration
30mCUDA in Your Python: Effective Parallel Programming on the GPU