NVIDIA scales financial clustering to 1M instruments across 64 GB200 GPUs

PyTorch · x · 2026-09-11

An NVIDIA Technical Blog post (shared by PyTorch) details AdaptGrow, a GPU-accelerated SymNMF solver for financial instrument clustering. A memory-efficient formulation cuts peak storage from 20n² to 4n² bytes, enabling factorization of 100,000 instruments on a single GB200; a distributed NCCL-based row-sharding scheme scales to 1M instruments across 64 GB200s on 16 nodes. The solver auto-configures from the eigenspectrum (full-batch AdaGrad vs block-stochastic SVRG), and ARI stability tracking with self-calibrated 3σ limits detects planted membership changes, while TPDM surfaces co-crash events correlation misses. Full reproducible notebook included.

Original post →

More from Infra

Infra channel →