New GPU solver speeds up ad matching 10x with near-linear scaling
This distributed PyTorch solver handles millions of users and thousands of items on multiple GPUs...
Deep Dive
Researchers present a distributed multi-GPU LP solver built natively in PyTorch for production decision systems like ad allocation. It uses column-sharded parallelism with fused Triton kernels, ridge-regularized LPs to improve stability, and an operator-centric programming model. On synthetic workloads, it achieves an order-of-magnitude wall-clock speedup over DuaLip-Scala and near-linear scaling (3.86x on 4 GPUs), enabling scale beyond existing GPU solvers.
Key Points
- Column-sharded parallelism with fused Triton kernels reduces per-iteration overhead and limits communication to item-level dual variable reduction.
- Ridge-regularized LPs provide explicit control over solution stability, a feature missing from cuPDLP and D-PDLP.
- Achieves 3.86x speedup on 4 GPUs and order-of-magnitude faster wall-clock times vs DuaLip-Scala on synthetic workloads.
Why It Matters
Enables larger, more stable production matching systems (ad allocation, recommendations) with fewer GPUs and simpler code changes.