Arteris proposes direct NoC-to-NoC die-to-die links for physical AI

Arteris outlined an approach where the network-on-chip spans chiplet boundaries directly instead of converting traffic into AXI transactions. This cuts overhead and latency when moving tensors and data streams between accelerators in automotive physical AI systems.

• NoC packets cross die boundaries natively, without AXI conversion
• Coherent protocols add up to 64 bytes of overhead per cache line
• Virtual channels separate control, sensor and AI inference traffic
• The approach targets terabyte-per-second automotive interconnect traffic