Mobility & logisticsResearch findings
HCRMap: Pressure-Aware Hot Expert Residency Mapping for 3.5D MoE Inference
Original published 2026.07.13 23:05 KSTAdded to Blackdeck 2026.11.17 06:17 KST
What has changed?
Dynamically manages hot expert residency across memory tiers to balance compute and communication load.
Key points
- HCRMap is proposed to address hot expert skew in 3.5D MoE chiplet systems by dynamically managing expert residency.
- It selects expert replicas based on hotness, loading cost, and runtime pressure to reduce communication and memory bottlenecks.
- End-to-end latency is reduced by up to 46.7% over PIMoE and 43.6% over Hydra in prefill and decode stages.
Connect it to your work
Proposes HCRMap, a framework that dynamically maps hot expert replicas to mitigate imbalance in 3.5D MoE inference systems.
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.