NAMD supports CPU-only optimizations for processors that implement particular AVX-512 vector instruction sets (avx512f and avx512dq). These instruction sets have been available for several years with Intel Xeon and more recently with AMD Zen4 processors. A vectorized kernel replaces the standard nonbonded force kernel, reimplementing the CUDA tiles algorithm for AVX-512 using the Intrinsics for Intel AVX-512 Instructions. Use of these optimizations is not compatible with GPU-accelerated builds.
The downloads page offers a multicore AVX-512 binary built using GCC 11.2.1 that is compatible with either Xeon or Zen4. Otherwise, build target Linux-AVX512-icx (using a recent Intel oneAPI toolkit) is recommended for Xeon and build target Linux-AVX512-g++ (using GCC version 11 or later) is recommended for Zen4. The AVX-512 tiles optimization is enabled by default. Since use of this optimized kernel is not compatible with all features, NAMD will automatically disable its use when an incompatibility is detected, falling back to the standard CPU code path. There is also a config file parameter to explicitly disable use of the tiles optimization.
Use the AVX-512 tiles optimization 