ILNumerics Accelerator (.NET) - Benchmarks and Artifacts 2026

This repository collects performance benchmarks for the ILNumerics Accelerator JIT compiler. This compiler is the driving force behind high-performance numerical array code optimizations for the ILNumerics numerical application development platform on .NET. The initial state of the repository (May 2026) represents artifacts for the 2026 paper "A Virtual Processor brings back the Free Lunch" published by ILNumerics GmbH, using ILNumerics version 7.4. It is subsequently extended with more and updated benchmarks as ILNumerics Accelerator evolves.

Each benchmark can be downloaded and reproduced locally. Docker images / build scripts are provided where suitable.

Content

Part 1 - Low Level Expressions Performance comparison of executing the array expression sum((m0 & (A << shift)) | (~m0 & B), dim: 1) with moderate array sizes on NumPy, FORTRAN and ILNumerics Accelerator (.NET).
Part 2a - Parallel.For vers. ILNumerics Accelerator Compares the performance of automatic parallelization by ILNumerics Accelerator (.NET) with the manual Parallel.For attempt on the array expression sum(abs(sin(A))).
Part 2b - ILNumerics Accelerator vers. Invalid Parallel.For Demonstrates strong-scaling capabilities by ILNumerics Accelerator (.NET) where Parallel.For fails: an order dependent loop is parallelized.
Part 3 - Strong Scaling the K-Means Algorithm` ILNumerics Accelerator (.NET) performs strong scaling on the iterative K-Means algorithm and beats FORTRAN by factor 3.