Intel® Graphics Technology for general purpose computing
summary
Claim: integrated Intel processor graphics can give up to 3× speedup but only for well‑matched codes and hardware; implementation uses OpenMP 4.0 or Intel pragmas and API with options for synchronous pragmas or asynchronous JIT‑style queues and explicit pin/share calls. Hardware specifics matter: Haswell maps to about 40 EUs with 7 hardware threads each (~280 threads), 4 KB per hardware thread register budget, 256 KB L3 per slice and optional 128 MB eDRAM, so collapse loops, vectorize (4‑byte floats/ints), tile to fit 4 KB, and watch OS/timeout and pointer/SVM limitations.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 0:54Technology overview
- 2:01Silicon architecture
- 3:00Performance gains
- 4:33Compiler methods
- 6:27Code parallelization
- 8:02Synchronous offloading
- 10:03Hardware specifics
- 16:16Programming models
- 18:15Implementation details
- 28:37OpenMP offload
- 32:08Async offload API
- 34:23Async example
- 38:49Best practices
- 42:38Performance tuning
- 48:20Matrix multiplication
- 52:35Compilation and tools
- 55:33Intel tools access
- 56:32Q&A session
related talks
31mGPU Performance Portability Using Standard C++ with SYCL
01h 07mAnother fundamental shift in Parallelism Paradigm?
01h 01mWant fast C++? Know your hardware!
52mAdvanced SIMD Algorithms in Pictures
01h 09mComputer Architecture, C++, and High Performance
01h 00mIntroduction to Hardware Efficiency in Cpp
01h 00mJust-in-Time Compilation
01h 04mCPU design effects