“A Fast Locality Simulator for GEMM Design-Space Exploration on Multi-Chiplet GPUs”, Euijun Chung, Hyesoon Kim, ModSim’26

  • Built an MI300X-like simulator that treats two XCDs, one IOD, and two HBMs as a single chiplet and models the amount of requests that miss within a chiplet.
  • Found that there’s no single best work/data scheduling strategy across tiled GEMMs. Then, used Claude to search the simulator’s design space for each GEMM’s scheduling configurations, which successfully improved L2 hit rate, cross-chiplet traffic, and energy efficiency.

“Enabling Efficient GPU Communication over Multiple NICs with FuseLink”, Kai Chen, OSDI’25
“TurboBus: Pooling PCIe Bandwidth for LLM Workloads via Scale-Up Fabrics”, Kai Chen, SIGCOMM’26

  • Same concept: GPUs relay over NVLink to borrow NICs (FuseLink) or PCIe bandwidths (TurboBus).