Abstract

Wafer-scale GPUs integrate many GPU modules (GPMs) on a single wafer, but their distributed organization makes the address translation non-uniform: on an L2 TLB miss, a page table walker may fetch a leaf-level PTE from a remote GPM across the on-wafer mesh. Across 16 irregular applications, ideal full replication reduces execution time by 54.9% but requires 36× of the storage of an unreplicated page table in a 6×6 mesh topology, equivalent to approximately 7.2% of the mapped memory footprint with 4KB pages. To this end, we propose Clover, a storage-efficient page-table replication mechanism that partitions the page table and applies a topology-aware k-hop placement. It allows every GPM to reach any partition within a small bounded number of hops while preserving conventional page-walk logic and keeping replicas coherent at low traffic. Clover reduces execution time by 44.9% on average while reducing aggregate page-table storage from 36× to 7.2× the unreplicated page-table size.