Intel has achieved a groundbreaking milestone in integrated photonics technology for high-speed data transmission, implementing the first fully integrated optical input/output (I/O) chiplet for scalable AI infrastructure.

Intel's Integrated Photonics Solutions (IPS) group announced on the 27th that it has demonstrated a cutting-edge technology that runs real-time data by co-packaging an industry-first fully integrated Optical Compute Interconnect (OCI) chiplet with an Intel CPU at the Optical Fiber Communication Conference 2024 (OFC 2024) held in San Diego, USA.

Intel’s OCI chiplet represents a quantum leap in high-bandwidth interconnect by enabling co-packaged optical I/O in emerging AI infrastructure for data center and high-performance computing (HPC) applications.

The first OCI chiplet to be unveiled is designed to support 64 channels of 32Gbps data transmission in each direction over optical fiber distances of up to 100m, and is expected to meet the demands of AI infrastructure for higher bandwidth, lower power consumption, and longer reach.

It also supports future scalability of CPU/GPU cluster connections and new compute architectures including consistent memory expansion and resource isolation.

The integrated OCI chiplet leverages Intel’s field-proven silicon photonics technology and integrates a silicon photonics integrated circuit (PIC) that includes an on-chip laser and optical amplifier with an electrical IC. The OCI chiplet implemented in OFC is packaged with Intel CPUs, but can also be integrated with next-generation CPUs, GPUs, IPUs, and other SOCs (systems-on-chips).

This first integrated OCI chiplet supports up to 4Tbps bidirectional data transfer and is compatible with peripheral component interconnect express (PCIe) Gen 5.

The demonstration demonstrated a real-time optical link showing transmitter (Tx) and receiver (Rx) connections between two CPU platforms over single-mode fiber (SMF) patch cords.

The CPU generated and measured the optical Bit Error Rate (BER), and the demo demonstrated a Tx optical spectrum using eight wavelengths spaced 200 GHz over a single fiber, along with a 32 Gbps Tx eye diagram demonstrating strong signal quality.

Current chiplets support 64 channels of 32 Gbps data in each direction for distances up to 100 m (though practical applications may be limited to tens of meters due to propagation delays) using eight fiber pairs, each carrying eight dense wavelength-division multiplexing (DWDM) wavelengths.

Additionally, the copackaged solution is highly energy efficient, consuming only 5 picojoules (pJ) per bit compared to approximately 15 picojoules (pJ) per bit for pluggable optical transceiver modules.

This level of ultra-high efficiency is critical for data centers and high-performance computing environments, and could help address the unsustainable power demands of AI.