How to run Neural Nets on GPUs'
summary
Shows actual code launching a neural net on a GPU and reports concrete tradeoffs: uses CUDA and cuDNN (and mentions cuBLAS/OpenCL), runs a Java DL4J MNIST Java DL4J example with batch size 100 getting about 69% accuracy on 1,000 examples, and emphasizes training speedups like SGD up to 40x faster. Notes practical limits: ~4 GB GPU memory, data transfer latency, mini‑batching and multi‑GPU distribution as real engineering concerns.
AI-written summary · about
chapters (38)
- 0:00Introduction
- 0:29Outline
- 1:12Neural Nets
- 4:59Why Neural Nets
- 6:16Personalization
- 7:43Computer Vision
- 9:40Training Time
- 10:16Graphics Processing Units
- 10:40CPU vs GPU
- 12:25GPU terminology
- 13:53Memory limits
- 14:38Moving memory
- 15:57How much memory
- 17:54Branching
- 18:22One Task
- 18:56Linear Calculation
- 19:20Linear Calculation Example
- 21:23Packages
- 21:55Config Files
- 22:56Code
- 23:32amnesty
- 23:52image classification
- 24:02structure
- 24:53class
- 25:14variables
- 25:36iterator
- 25:58live coding
- 26:51seed
- 27:30iterations
- 28:54layers
- 29:37activation
- 30:25output
- 31:08model
- 32:09backend
- 33:22Troubleshooting
- 35:06Results
- 35:17Examples
- 36:30SkyyMind
