AI accelerators — tensor cores, systolic arrays, spatial architectures — are designed for matrix multiplications. Scientific computing also uses matrix operations but requires different numerical patterns: sparse linear algebra, iterative solvers, stencil computations. Can AI hardware do general numerics?
Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2603.23343. The paper implements three numerical kernels and a preconditioned conjugate gradient solver on Tenstorrent's Wormhole and reports that a single Wormhole die runs within 4.3 times (16-bit) and 8.8 times (32-bit) of an Nvidia H100, despite an older process and much lower memory bandwidth, concluding that AI accelerators merit consideration and that more work is needed. The essay says spatial architectures bridge the gap better than GPU architectures; the paper's measurements show the Wormhole solver slower than the H100, and the claim about better suitability for stencil and iterative codes is not in the abstract.
The authors (arXiv:2603.23343) implement numerical kernels and a conjugate gradient solver on Tenstorrent's Wormhole spatial architecture. The spatial design — where data flows between processing elements without centralized control — maps naturally onto stencil operations and iterative algorithms where each element communicates with neighbors.
The through-claim: AI accelerators and scientific computing share the same fundamental operation (matrix computation) but diverge in access patterns and control flow. Spatial architectures bridge this gap better than GPU architectures because the spatial data flow matches the locality structure of scientific computations. The hardware designed for neural network layers turns out to be well-suited for PDE solvers — not because the problems are similar, but because both are local computations over regular grids.