Intel Begins Working On "Knights Mill" Support For LLVM/Clang

Intel compiler engineers have begun mainlining "Knights Mill" enablement within the LLVM compiler stack.

Knights Mill is the codename for an upcoming Xeon Phi expected for release later this quarter. Details on Knights Mill are relatively light but it will cater to deep learning / AI use-cases and more efficient than Knights Landing (KNL).

Intel has previously said Knights Mill is capable of twice the performance of Knights Landing for floating point operations per cycle and there are also new/optimized instructions for 8-bit and 16-bit arithmetic.

In deep learning, Knights Mill should deliver 4x the performance of KNL. Other known details are six-channel DDR4-2400 support, 16GB of MCDRAM, quad FMA support, and Knights Landing will continue to be available for HPC customers while Knights Mill is centered on deep learning / AI markets.

Hitting LLVM today in Git is the initial enablement as "KNM" CPU target for Knights Mill. Currently it's an alias of Knights Landing, but the patch does also confirm AVX5124FMAPS and AVX5124VNNIW will come for this Xeon Phi. There's also a similar patch to Clang. Expect more Intel "KNM" patches to come soon and optimizations.