Ayar Labs TeraPHY Optical I/O Chiplet

TeraPHY is Ayar Labs' monolithic electronic-photonic optical I/O chiplet fabricated on standard GlobalFoundries CMOS silicon photonics processes. Designed to be co-packaged or 2.5D/3D integrated with GPUs, TPUs, and custom AI accelerators via standard UCIe die-to-die interfaces.\n\nINFERENCE RELEVANCE: Large-scale LLM inference (particularly autoregressive decode with large batch sizes, or massive MoE models requiring all-to-all expert routing across thousands of chips) is throttled by copper trace reach and serializer/deserializer (SerDes) power. Copper NVLink and PCIe direct-attach cables are physically limited to ~1 meter without bulky, power-hungry active electrical cables (AEC) or retimers. TeraPHY delivers 10 ns latency per chiplet (excluding optical flight time in fiber) and sub-pJ/bit energy consumption, allowing multi-rack clusters to scale tensor-parallel and pipeline-parallel domains across distances of millimeters to kilometers at scale-up latencies. By decoupling physical proximity from scale-up domain boundaries, optical I/O enables distributed LLM inference clusters to treat rack-scale or multi-rack accelerator pools as a single coherent memory and compute fabric without the latency overhead of conventional multi-hop Ethernet/InfiniBand networks.