Chips & computingResearch findings
Mozart: A Framework for Composable Custom ASIC Design
Original published 2025.10.10 09:07 KSTAdded to Blackdeck 2026.11.20 04:21 KST
What has changed?
Mozart enables operator-level disaggregation to design cost-effective, heterogeneous ASICs tailored to neural network patterns.
Key points
- Mozart systematically designs custom ASICs by leveraging operator-level heterogeneity and chiplet diversity.
- It enables efficient, heterogeneous acceleration through disaggregation and tensor fusion at the operator level.
- Demonstrates up to 78.8% reduction in energy-delay-cost product across datacenter LLM serving, speculative decoding, and autonomous vehicle
Connect it to your work
With just 8 strategically selected chiplets, composite ASICs achieve significant energy and cost efficiency improvements over traditional accelerators.
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.