Intel Math Kernel Library Perspectives & Latest Advances
summary
Intel MKL now targets Xeon Phi with three concrete execution models: automatic offload where the MKL runtime transparently splits work between host and coprocessor (applies to a subset like level‑3 BLAS GEMM, TRSM/TRMM and LAPACK LU/QR/Cholesky); compiler‑assisted offload via pragmas for explicit data persistence and reduced transfers; and native execution by building and running on the card. MKL adds conditional numerical reproducibility to force run‑to‑run, inter‑CPU and cross‑vendor identical results at the cost of some performance (compiler FP model “precise” shows ~12–15% slowdown in benchmarks). Tools: link‑line advisor and MKL runtime controls to tune work division.
AI-written summary · about
chapters (6)
- 0:00Introduction to MKL
- 2:27MKL and Xeon Phi support
- 10:01Execution models for Xeon Phi
- 23:12Tools and resources
- 27:48Numerical reproducibility
- 42:42Summary and conclusion
related talks
30mOptimizing vLLM for Intel CPUs and XPUs
30mAI: unleashing the next wave
01h 00mThe Blaze High Performance Math Library
11mEnabling Composition of Kernels and Compilers
14mMetaKernel: Enabling Efficient Encrypted Neural Network Inference Through Unified MVM(…)
41mML applications in Linux kernel
40mHidden GEMMs: How Optimized Math Libraries Work
08mGPU Programming with Triton Kernels