Numerical systems for accelerated science.

A systems research group, not a simulation vendor.

We do not claim domain ownership over CFD, chemistry, climate, or materials science. We build reusable numerical infrastructure underneath those applications and qualify how that infrastructure behaves on modern accelerators.

Methods, kernels, and evidence.

• Mixed-precision and precision-recovery methods.
• Dense, sparse, spectral, and solver kernel implementations.
• ROCm/HIP and CUDA-class accelerator qualification.
• Multi-GPU communication and topology characterization.
• Machine-readable experiment plans, run receipts, and negative-result records.

Teams whose science is bottlenecked by the machine.

Accelerator vendors, OEMs, scientific-software teams, HPC centers, research labs, and AI-for-science groups that need to turn new hardware into reliable numerical capability without sacrificing the correctness contract.

What is live today

ShardNet currently publishes algorithmic experiments and implementation self-tests. Physical AMD Instinct qualification is explicitly marked pending until the corresponding hardware run has been completed and retained with its full environment and measurement receipt.