Suanmiao Technology’s Domestic 3D TokenPU Chip Completes Tapeout with 16 TB/s Bandwidth

Suanmiao Technology has completed tapeout of its domestically developed 3D TokenPU chip. The chip targets large model inference workloads. It uses a 3D hybrid stacking architecture to achieve 16 TB/s bandwidth.

3D hybrid stacking enables 16 TB/s bandwidth

The 3D TokenPU is an AI inference processor designed and manufactured entirely within China. The company claims the chip breaks foreign monopoly in this segment.

The chip's 3D hybrid stacking architecture enables 16 TB/s bandwidth. This high bandwidth reduces data transfer loss and improves inference efficiency for large models, according to Suanmiao Technology.

The entire process from design to manufacturing is domestic. This marks a step toward self-sufficiency in advanced AI hardware.