Mobility & logisticsResearch findings
A Method for Accelerating Low-Batch MoE Inference Using Multi-Chiplet Architecture
Original published 2026.03.29 19:44 KSTAdded to Blackdeck 2026.11.18 03:13 KST
What has changed?
A dynamic expert trajectory scheduling method that balances workload across multi-chiplet systems for low-batch MoE inference
Key points
- A dynamic expert trajectory scheduling method is proposed for low-batch MoE inference on multi-chiplet accelerators
- FSE-DP achieves 1.22 to 2.00x speedup over state-of-the-art and reduces on-chip memory usage by up to 78.8%
- Fine-grained expert streams are orchestrated across high-bandwidth D2D links to balance load and improve computation-communication overlap
Connect it to your work
FSE-DP enables adaptive computation-communication overlap and load balancing via fine-grained expert streams over high-bandwidth D2D links
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.