#13 · Tools & infrastructure

GPU-batched linear algebra speeds genome-scale metabolic and structural computations

The dominant thread is reformulating repeated, independent computations as batched or GPU-parallel linear algebra. manylp solves batches of bounds-parametric LP instances for genome-scale dynamic flux balance analysis, reusing cached bases to certify unique vertex fluxes on CPU or GPU while maintaining exactness. gpuGEM applies NVIDIA cuOpt through an open-source Python library to GPU-based linear optimisation of genome-scale metabolic models, reporting speed-ups. In HADDOCK3, pairwise RMSD is recast as batched SVD and FCC clustering as binary matrix multiplication, yielding 6.41x on MPS and up to 12.2x on CUDA without structural regression. What is new across these items is the pairing of throughput with a stated guarantee: exactness for manylp and structural fidelity for the HADDOCK3 work. The available gpuGEM summary does not state a comparable guarantee alongside its reported speed-ups.