Arteris is proposing a multi-die interconnect model that extends the network-on-chip itself across chiplet boundaries. Instead of terminating the NoCnetwork on chipA NoC is the on-chip communication fabric used to move data between functional blocks. at each die and translating traffic into cache-line transactions or AXIAdvanced eXtensible InterfaceAXI is an AMBA interface used widely for high-performance on-chip data movement. transfers, the architecture keeps the native packet format intact over the die-to-die link. The goal is a stable, implementation-independent interface between chiplets.
The motivation comes from the traffic profile of physical AIartificial intelligence systems. According to the source text, autonomous vehicles, industrial robots, humanoid robots, drones, and intelligent manufacturing platforms increasingly move large streams of feature maps, tensors, point clouds, video frames, and intermediate inference results between specialized accelerators. These transfers are described as continuous, deterministic, and often far larger than the small transactions for which coherent interconnects were originally optimized.
Arteris contrasts this with conventional coherent die-to-die protocols, which are well suited to memory-centric processor workloads but less efficient for accelerator-to-accelerator traffic. Repeating cache-line metadata across many 64-byte transactions adds overhead when payloads extend into hundreds of kilobytes or megabytes. The same critique is applied to non-coherent approaches that extend AXIAdvanced eXtensible InterfaceAXI is an AMBA interface used widely for high-performance on-chip data movement. across chiplets: they may connect peripherals well, but they still force packet-to-transaction conversion on both sides of the boundary.
The proposed answer is an invariant die-to-die NoCnetwork on chipA NoC is the on-chip communication fabric used to move data between functional blocks. interface. By carrying protocol-independent NoCnetwork on chipA NoC is the on-chip communication fabric used to move data between functional blocks. packets directly across the interface, each chiplet can keep evolving internally while the external architectural contract remains stable. The source says this approach supports virtual channels, can be sized for future bandwidth needs, and extends NoCnetwork on chipA NoC is the on-chip communication fabric used to move data between functional blocks. routing across dies without repeated reconstruction of packet flows.
For next-generation automotive AIartificial intelligence platforms, the claimed benefits are lower latency, reduced buffering, better bandwidth utilization, and lower energy per transferred bit. The article frames this as a way to scale multi-chiplet systems without exposing internal NoCnetwork on chipA NoC is the on-chip communication fabric used to move data between functional blocks. changes to the interconnect contract that links independent dies.