A method that allows sequential programs to be executed efficiently and automatically on modern, parallel computer hardware

ILNumerics presents a solution to a global problem:
Automatic optimal use of modern hardware by complex programs
Berlin, 10.11.2025
The Berlin-based technology start-up ILNumerics GmbH has so far been known for its innovative software solutions in math-heavy, technical application development in the rather small .NET community. But now it is causing a stir in the software world. The small company seems to have succeeded in what entire development departments and universities have been looking for for more than a decade: a method that allows sequential programs to be executed efficiently and automatically on modern, parallel computer hardware.
In doing so, ILNumerics addresses a central problem of modern software development: despite enormous development efforts, it is still hardly possible to efficiently exploit the parallel resources of modern hardware. The so-called "parallel programming gap" means that many technical programs can only make use of small shares (3%... 30%) of modern hardware resources. The global consequences are massive energy and time losses as well as high costs due to inefficient resource use.
At the heart of the new technology is a new type of execution scheme that equips programs with a form of artificial intelligence: ILNumerics gives the elementary building blocks of technical programs - the so-called array operations - just enough intelligence to decide autonomously on which hardware, when and how to execute themselves. This autonomy dissolves the original sequential structure of a program - only to have it merge again in parallel at runtime when all system state is known. ILNumerics can skip rigid graph optimizations and makes use of the parallel potential throughout large program parts - across functional boundaries.
From an originally linear program, the computational result emerges as a self-organizing structure of hundreds to thousands of parallel executions. Each of these paths makes local decisions in real time that no human developer could ever anticipate at this level of granularity or speed: which resource (CPU, GPU, etc.), which implementation (low-level code) currently promises the fastest completion and which operations should run concurrently.
The result: programs that automatically execute optimally - hardware-agnostic and autonomous. Without any time-consuming manual runtime planning complex programs are created much faster and deliver significantly faster results. The technology could initiate a radical change in software development, as it automates entire process chains to optimize speed. Like neurons in the brain, the overall behavior of the program is not created by central control (programmer / scheduler), but emerges collectively - through the cooperation of many, locally meaningful, parallel, autonomous decisions.
The software offers an array language compatible with Matlab and NumPy and is available under commercial and non-commercial licenses. It is aimed at application developers in technical domains who work with complex numerical algorithms and array data of any size. In benchmarks, the company presents speed-ups of 3... 60 times compared to traditional technologies, such as NumPy or FORTRAN, as well as to manually optimized, native codes.
(See: https://ilnumerics.net/speed-comparison.html).
The company is currently preparing strategic collaborations with leading software, cloud, and IT companies to integrate the autonomous optimization technology into cloud and other computing environments. This opens the perspective to a new generation of self-optimizing software - from laptops to supercomputers.

Haymo Kutschbach (Founder, CEO)
Bessemerstr. 51, 12103 Berlin, Germany

The Berlin-based technology start-up ILNumerics GmbH has so far been known for its innovative software solutions in math-heavy, technical application development in the rather small .NET community. But now it is causing a stir in the software world. The small company seems to have succeeded in what entire development departments and universities have been looking for for more than a decade: a method that allows sequential programs to be executed efficiently and automatically on modern, parallel computer hardware.

This release was published on openPR.