← Back to the feed
AI & softwareResearch findings

Compass: Co-Exploration Framework for Dynamic LLM Workloads in Heterogeneous Multi-Chiplet Accelerators

Original published 2025.12.06 04:03 KSTAdded to Blackdeck 2026.11.18 16:56 KST

What has changed?

A computation execution graph-based mapping scheme decouples micro-batch and layer dimensions, enabling fine-grained control on heterogeneous chiplets

Key points

  1. Large language models bring massive computational demands requiring multi-chiplet accelerators
  2. Existing design space exploration focuses on traditional workloads, neglecting dynamic LLM behavior
  3. Compass jointly optimizes mapping and hardware design for dynamic LLM inference workloads

Connect it to your work

Proposes a co-exploration framework that jointly optimizes mapping strategies and hardware design for dynamic LLM inference workloads

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.

  1. arxivCompass: Co-Exploration of Mapping and Hardware for Heterogeneous Multi-Chiplet Accelerators Targeting LLM Inference Service Workloads