ILNumerics Accelerator – Demos & Benchmarks

Benchmarks on GitHub

The ILNumerics/decentralized-array-execution-artifacts2026 artifacts repository on GitHub hosts benchmarks, originating from our 2026' paper: 'A Virtual Processor brings back the Free Lunch' (pdf). We will try to keep it updated with new releases. The benchmarks are designed to make reproduction really easy. Go ahead and reproduce them with your own hardware and let us know your results! 

Demo apps

Some of the artefacts on GitHub stem from the following benchmarks. We use them to demonstrate first-touch experience with the Accelerator and to explain fundamental concepts associated with it. Topics covered include: verifying accelerated code, verifying acceleration factors, configuring acceleration and making sense of measurement results. 

• Here, we compare the execution speed achieved by NumPy, Numba, Fortran and ILNumerics for the low-level expression $sum((m0 & (A << shift)) | (~m0 & B), dim: 1)$    
• A number of getting started guides deep-dive into practical aspects of automatic code acceleration. 
  • Part I investigates the impact on a simple reduction operation 
  • Part II investigates optimizations applied by the compiler to a larger function context (kmeans)
  • Part III computes a large number of FFTs faster than ... Intel's MKL
  • Part IV examines the advantages of array instruction parallelism and compares them with legacy methods