Machine Learning Frameworks Interoperability
summary
Pipeline reads CSV straight into cuDF on the GPU, uses CuPy FFT and a Numba/CUDA kernel to segment and embed heartbeats, then trains a PyTorch VAE on 100k length‑250 vectors and visualizes with cuXfilter; Jupyter notebooks reproduce it. Claim: ASCII parsing and layout conversion costs ~70% of runtime; solution: RMM shared memory, zero‑copy via DLPack/CUDA array interface, and RDMA/GPU Direct Storage to avoid CPU staging and cut multi‑hop copies by about a factor of five — the toy pipeline runs end to end.
AI-written summary · about
chapters (6)
- 0:00Framework interoperability
- 2:54Heartbeat segmentation pipeline
- 10:17Memory management across frameworks
- 13:06Data conversion bottlenecks
- 18:41Hardware-accelerated RDMA copying
- 23:21Summary
related talks
39mAccelerating Data Science with RAPIDS
41mAccelerating Data Science with RAPIDS
29mRay Data on GPU - Breaking Preprocessing Bottleneck for Model Training
01h 08mGPU Data Science with Rapids
34mSpeeding up Machine Learning tasks using GPUs in Python
01h 22mHigh-Performance Data Science at Scale with RAPIDS, Dask, and GPUs
39mAccelerated Data Science Libraries
01h 12mGPU-Accelerated Data Analytics in Python