XAMBA: Enabling Efficient State Space Models on Resource-Constrained Neural Processing Units

Arghadip Das, Arnab Raha, Shamik Kundu, Soumendu Kumar Ghosh, Deepak Mathaikutty, and Vijay Raghunathan · 2025

Read XAMBA: Enabling Efficient State Space Models on Resource-Constrained Neural Processing Units by Arghadip Das, Arnab Raha, Shamik Kundu, Soumendu Kumar Ghosh, Deepak Mathaikutty, and Vijay Raghunathan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-06924.txt.

Description

Arghadip Das and co-authors bring state space models to the small neural chips built into AI laptops. These models rival transformers on long texts but sip far less compute. The catch is that each new network breed tends to demand its own fresh chip.

Their XAMBA frame instead tunes the models for chips already on sale. First it gets the models running, then it speeds them up, then it trades a little accuracy for more speed. Two tricks swap step-by-step sums for matrix math, which the chips love. A third trick swaps costly curve functions for straight-line chunks, cutting wait with little loss.

Catalog record

CREATORArghadip Das, Arnab Raha, Shamik Kundu, Soumendu Kumar Ghosh, Deepak Mathaikutty, and Vijay Raghunathan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-06924

Browse the full library · About the Tower of Babel