NumPy 2.0 was released on Sunday that's been in the making for the past year and their first major release since 2006. While it comes with API/ABI breakage, NumPy 2.0 delivers new features and performance improvements.

Interesting on the performance front is that NumPy 2.0's sorting functions like sort / argsort / partition / argpartition are much faster than before. NumPy is now leveraging Intel's x86-simd-sort library and Google's Highway for achieving "large speedups" with the new NumPy version.

On Phoronix I've previously covered the massive performance potential of [1]x86-simd-sort with much faster sorting thanks to [2]leveraging AVX-512 and other modern ISA features. Google's Highway also helps with performance-portable SIMD with run-time dispatch. Highway helps with not only targeting AVX2 and AVX-512 but many processor family sub-targets, RISC-V RVV, ARM NEON, and other processors.

NumPy 2.0 also brings opt_func_info as a new tracing and introspection API for probing hardware-specific kernels available and dispatching to them. NumPy 2.0 additionally brings a new "StringDType" variable-length string dtype, many string improvements, support for float32 and longdouble in NumPy FFT functions, and other C and Python API additions.

Downloads and more details on NumPy 2.0 via [3]NumPy.org .



[1] https://www.phoronix.com/search/x86-simd-sort

[2] https://www.phoronix.com/news/Intel-AVX-512-Quicksort-Numpy

[3] https://numpy.org/news/