TIDE-MC: New GPU framework crushes memory limits with 11,647x speedup

Matrix completion is critical for large-scale recommendation systems and scientific computing, but existing GPU solvers often assume the observed matrix or dense factors fit in device memory—a promise that fails on real workloads, causing out-of-memory errors or severe PCIe bottlenecks. TIDE-MC, presented by a team from Nanjing University and collaborators, breaks this barrier with a Two-Sided Interpolative Decomposition (TSID) strategy. Instead of processing the full matrix, TIDE-MC samples a small template submatrix as an anchor, then reconstructs the entire low-rank matrix from that template plus active data chunks. This design keeps memory usage proportional to the template and chunk size, not the full dataset.

TIDE-MC executes in two stages. First, a conflict-free synchronization engine recovers the template via parallel factorization and hierarchical gradient aggregation. Second, a chunked reconstruction pipeline extends the template to the rest of the matrix, overlapping PCIe transfers with GPU computation to hide I/O latency. An asymmetric gradient-clipping scheme stabilizes mixed-precision Tensor Core training, enabling healthy convergence. In tests against state-of-the-art baselines on 15 benchmarks, TIDE-MC completed workloads that crashed others, delivering speedups up to 11,647x, peak memory reductions up to 8.5x, and reconstruction error improvements up to 99.7%. This proves that template-anchored decomposition can scale matrix completion far beyond what device memory alone would allow.