Guillaume Verdon introduces Z1T, a family of transformer-like models built for Extropic's sparse probabilistic Z1 hardware. Instead of forcing dense matrix operations onto the chip, the team modifies model connectivity, activations and attention so the algorithms match the hardware's locally connected sparse graph.
Verdon says the chip uses physical transistor noise to represent probabilistic values, with repeated sampling providing effective precision. He reports roughly 1,000 times better efficiency for layers moved onto Z1 and more than 100 times better efficiency for the mixed workload once the remaining conventional accelerator work is included.
The current open model is around GPT-2 scale, but Extropic is publishing weights and the training recipe so researchers can explore larger sparse variants. Verdon argues that scaling connectivity alongside compute could open a new design space for inference systems constrained by data-center power.
Watch on YouTube
