← Back to the feed
AI & softwareResearch findings

Hemlet: A Hybrid Chiplet Architecture for Vision Transformer Acceleration

Original published 2025.11.19 21:49 KSTAdded to Blackdeck 2026.11.19 00:33 KST

What has changed?

A hybrid chiplet architecture integrating analog, digital CIM, and intermediate data processing units for flexible resource scaling.

Key points

  1. Addresses high memory and computational demands of Vision Transformers (ViTs).
  2. Enables flexible resource scaling through heterogeneous CIM and IDP chiplets.
  3. Achieves throughput up to 9.56 TOPS with energy efficiency of 4.98 TOPS/W.

Connect it to your work

Group-level parallelism and system-level dataflow optimization reduce communication overhead and improve throughput.

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. arxivHemlet: A Heterogeneous Compute-in-Memory Chiplet Architecture for Vision Transformers with Group-Level Parallelism