📢Out now: a study by @nathanleroux.bsky.social and colleagues from @fz-juelich.de demonstrates that using gain cell-based in-memory computing can accelerate attention in LLMs, training a 1.5B model with up to 70,000x energy savings and 100× speedup over GPUs. www.nature.com/articles/s43...

I’m proud to share our new paper in @natcomputsci.nature.com : www.nature.com/articles/s43.... We show that attention in LLMs can be accelerated with analog in-memory computing using Gain Cell circuits. Simuating a 1.5B-parameter model we achieve up to 70 000× lower energy and 100× speedup vs GPUs.