Bencheikh Wadjih

Leads SpiNNcloud’s HPC application engineering team (6 people) and owns the roadmap for sparse workloads on SpiNNaker2’s 5-million-core platform. Drives production HPC projects with partners including Sandia National Laboratories and Siemens, covering sparse solver and FEM workloads. Improves the software stack from host APIs to on-chip kernels, including Kokkos integration in collaboration with the Kokkos team, to reach HPC maturity and strong performance/efficiency for sparse computation.

Led hardware–algorithm co-design efforts to prepare SpiNNaker2 for HPC and sparse workloads, collaborating directly with chip architects on next-generation designs. Deployed a 1B-parameter Mixture of Experts LLM for on-chip inference (externally trained), managing expert routing and dispatch under tight memory/bandwidth budgets. Developed a general-purpose sparse linear algebra computation stack, including host-side Python APIs, on-chip C/C++ kernels, runtime support, profiling, and benchmarking, and demonstrated speed/energy improvements using a conjugate gradient solver for large Poisson problems.

Implemented memory-efficient spiking neural network (SNN) training on Graphcore IPU using surrogate gradients and online learning rules. Developed gradient checkpointing for sparse and recurrent architectures using off-chip memory to extend trainable model capacity (arXiv:2412.11810). Contributed to SNNAX, an open-source JAX-based spiking neural network framework.