The German company SpiNNcloud is now called Sparxity. The renamed company will supply a sparse-native rack system worth several million to the Jülich Supercomputing Center (JSC). That system, JUBIS, will run sparse ML models, HPC simulations, and quantum emulation.
Sparxity announced the deal today. The new JUBIS system is what’s known as an event-based system; it performs calculations only when a change occurs. The promised benefits of this “sparsity” approach are substantial. JUBIS is expected to execute sparse solvers 10 times faster and run event-based ML workloads using 18 times less energy than previous systems. For optimization problems, it should arrive at an acceptable solution 20 times faster.
JSC is a logical location for the computing power. It also houses JUPITER, Europe’s first exascale supercomputer. Its Booster partition achieved approximately 63 gigaflops per watt and topped the Green500 list. The JUPITER AI Factory, one of 19 European AI factories and the only one located near an exascale machine, also operates there.
From neuromorphic to sparse-native
The name change reflects a shift in focus. Whereas the company previously focused on developing a neuromorphic chip architecture based on technology developed at TU Dresden, the emphasis is now on sparse-native computing power.
The focus is now shifting entirely to sparse-native computing. “For years, model builders have had to design around the limitations of dense hardware,” says CEO Hector Gonzalez. Sparse computing reduces both the amount of computation and the amount of data that needs to be moved. According to Sparxity, AI models can be “sparser” than they are now, but the hardware is designed for “dense” computations.
Opportunity for research
JSC sees opportunities primarily in research. “Sparse machine learning is one of the greatest opportunities to fundamentally improve the efficiency of AI,” says Guido Trensch of the Simulation and Data Lab Neuroscience. Sparxity plans to make JUBIS more widely available for HPC and AI applications in the future.