Extropic Z1T (extropic.ai)

Extropic has unveiled the Z1T, a new architecture designed to enhance the energy efficiency of transformer models through a novel probabilistic sub-threshold CMOS chip. As the demand for transformer-based AI workloads escalates, leading to increased financial and energy costs, the Z1T aims to address these challenges by leveraging a sparse physical connectivity model. This allows for operations that are optimized for next-gen in-memory computing hardware, moving away from traditional dense matrix multiplication primarily optimized for GPUs. The Z1 chip employs probabilistic bits (pbits) to create a graphical model that enables efficient computation for deep learning tasks. By combining the Z1T with FPGA co-processors, the architecture can achieve a significant performance boost while reducing energy consumption—potentially providing up to three orders of magnitude more energetic efficiency compared to GPUs for specific tasks. This initial exploration into sparse neural networks and hardware co-design heralds a significant step towards a new era of AI models, one that aligns with both hardware capabilities and algorithmic advancements. Extropic is also open-sourcing key components of this project, fostering collaborative development in the AI community.