Energy & materialsResearch findings
LEXI: Lossless Exponent Compression for Efficient Inter-Chiplet Communication
Original published 2026.03.17 02:48 KSTAdded to Blackdeck 2026.11.18 07:23 KST
What has changed?
Demonstrated high compressibility in floating-point exponents and proposed a lossless compression scheme based on Huffman coding.
Key points
- Floating-point exponent entropy is low, indicating high compressibility.
- LEXI enables real-time compression and decompression during inter-chiplet communication.
- Reduces communication and inference latency by 33–45% and 30–35% respectively on Jamba, Zamba, and Qwen LLMs.
Connect it to your work
Achieves 33–45% reduction in inter-chiplet communication and 30–35% end-to-end inference latency on modern LLMs.
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.