#decadeai

First Integrated Optical I/O Chiplet

With integrated photonics technology, Intel Corporation has made significant progress towards high-speed data transmission. The first-ever fully integrated optical computing interconnect (OCI) chiplet, co-packaged with an Intel CPU and executing real data, was showcased by Intel’s Integrated Photonics Solutions (IPS) Group at the Optical Fibre Communication Conference (OFC) 2024. This chiplet is the most sophisticated in the industry. By enabling co-packaged optical input/output (I/O) in developing  AI infrastructure for data centres and high performance computing (HPC) applications, Intel’s OCI chiplet marks a significant advancement in high-bandwidth connection.

What It Does

This is the first OCI chiplet, intended to meet the increasing demands of  AI infrastructure for greater bandwidth, lower power consumption, and longer reach. It can support 64 channels of 32 gigabits per second (Gbps) data transmission in each direction on up to 100 metres of fibre optics. It makes it possible for CPU/GPU cluster connectivity to grow in the future and for innovative compute designs like resource disaggregation and coherent memory extension.

Why It Matters

Large language models (LLM) and generative  AI are two recent advancements that are speeding up the global deployment of AI-based applications. Machine learning (ML) models that are larger and more effective will be essential in meeting the new demands of workloads involving AI acceleration. Future  AI computing platforms will need to be scaled, which will require exponential expansion in I/O bandwidth and longer reach to support larger CPU/GPU/IPU clusters and architectures with more effective resource utilisation, like memory pooling and xPU disaggregation.

High bandwidth density and low power consumption are supported via electrical I/O, or copper trace connectivity, although its reach is limited to one metre or less. When employed in data centres and early  AI clusters, pluggable optical transceiver modules can expand reach at power and cost levels that are unsustainable for the scalability demands of AI workloads. AI/ML infrastructure scalability calls for co-packaged xPU optical I/O that can enable greater bandwidths with better power efficiency, longer reach, and low latency.

Electrical I/O

To use an analogy, switching from horse-drawn carriages, which had a limited capacity and range, to cars and trucks, which can transport much bigger amounts of products over much longer distances, is analogous to replacing electrical I/O with optical I/O in CPUs and GPUs to convey data. Optical I/O solutions such as Intel’s OCI chiplet could offer this kind of enhanced performance and energy efficiency to  AI scalability.

How It Works

The fully integrated OCI chiplet combines an electrical integrated circuit (IC) with a silicon photonics integrated circuit (PIC), which incorporates on-chip lasers and optical amplifiers, by utilising Intel’s field-proven silicon photonics technology. Although the OCI chiplet showcased at OFC was co-packaged with an Intel CPU, it can be combined with different system-on-chips (SoCs), GPUs, IPUs, and next-generation CPUs.

This initial OCI version is compatible with PCIe Gen5 and provides bidirectional data transmission rates of up to 4 terabits per second (Tbps). A transmitter (Tx) and receiver (Rx) connection between two CPU platforms via a single-mode fibre (SMF) patch cord is shown in the live optical link demonstration. The demonstration shows the Tx optical spectrum with 8 wavelengths at 200 gigahertz (GHz) spacing on a single fibre, along with a 32 Gbps Tx eye diagram demonstrating strong signal quality. The CPUs generated and tested the optical Bit Error Rate (BER).

The current chiplet uses eight fibre pairs, each carrying eight dense wavelength division multiplexing (DWDM) wavelengths, to provide 64 channels of 32 Gbps data in each direction up to 100 metres (though actual implementations may be limited to tens of metres due to time-of-flight latency). In addition to being incredibly energy-efficient, the co-packaged solution uses only 5 pico-Joules (pJ) per bit, as opposed to around 15 pJ/bit for pluggable optical transceiver modules.  AI’s unsustainable power requirements may be addressed with the help of this level of hyper-efficiency, which is essential for data centres and high-performance computing settings.

Concerning Intel’s Preeminence in Silicon Photonics

With over 25 years of in-house research from Intel Labs, the company that invented integrated photonics, Intel is a market leader in silicon photonics. The first business to create and supply industry-leading dependability silicon photonics-based connectivity solutions in large quantities to major cloud service providers was Intel.

The primary point of differentiation for Intel is their unmatched integration of direct and hybrid laser-on-wafer technologies, which result in reduced costs and increased reliability. Intel is able to preserve efficiency while delivering higher performance thanks to this innovative method. With over 8 million PICs and over 32 million integrated on-chip lasers shipped, Intel’s reliable, high-volume platform has a laser failures-in-time (FIT) rate of less than 0.1, which is a commonly used reliability metric that shows failure rates and the frequency of failures.

For use in 100, 200, and 400 Gbps applications, these PICs were installed in big data centre networks at prominent hyperscale cloud service providers in the form of pluggable transceiver modules. In development are next generation 200G/lane PICs to handle 800 Gbps and 1.6 Tbps applications that are only starting to gain traction.

Additionally, Intel is introducing a new fab process node for silicon photonics that offers significantly better economics, higher density, better coupling, and state-of-the-art (SOA) device performance. Intel keeps improving SOA performance, cost (more than 40% reduction in die size), power (more than 15% reduction), and on-chip laser performance.

What’s Next

This OCI chiplet from Intel is a prototype. Intel is collaborating with a small number of clients to co-package OCI as an optical I/O solution with their SoCs.

The OCI chiplet from Intel is a significant advancement in high-speed data transfer. Intel continues to be at the forefront of innovation and is influencing the future of connectivity as the  AI infrastructure landscape changes.