What Python Developers Need to Know About Hardware: A Practical Guide...
summary
Claims: GPU JIT compilation causes first-run stalls but caching avoids repeated compile latency; a CUDA kernel launch is thousands of times slower than a C call so kernel fusion is essential; GPUs run asynchronous streams so hiding launch latency and using multiple streams improves throughput; memory numbers are stark (CPU ~100 GB/s vs GPUs up to 8 TB/s) and massive thread counts (hundreds of thousands) drive throughput; tools named include CuPy, PyTorch, TensorFlow, JAX, Numba, CUDA and options for run-to-run determinism.
AI-written summary · about
chapters (8)
- 0:00GPUs for Python developers
- 2:53GPU compilation complexity
- 7:29Kernel launch latency
- 13:12CPU-GPU memory transfers
- 17:22GPUs maximize throughput
- 23:44Parallelism vs CPUs
- 26:51GPU pitfalls
- 27:33Q&A
related talks
01h 29mGPU development with Python 101
33mWhen GPUs Make Python Slower: Patterns and Pitfalls
02h 55mGPU Accelerated Python w: CUDA
48mGPU Programming in Pure Python
29mA Beginners Guide to GPUs for Pythonistas
30mUnderstanding GPUs
50mFrom Magic to Mechanics: Demystifying GPU Programming
01h 06mFoundation of GPU Programming