“We demonstrate that MatMul operations can be eliminated from LLMs while maintaining strong performance, even at billion-parameter scales.”
The billion-parameter scale tops out at 2.7B, a rounding error next to the models people pay for. Memory drops 61% in training and over 10x at inference. The 10x energy win comes from neuromorphic chips measured against edge GPUs, not a datacenter. “The gap narrows as size increases” is a forecast drawn from the small end of the curve.