Intel showcases its first fully integrated Optical I/O Chiplet

Another major breakthrough in the AI field? The media reported that Intel has launched its first OCI chip, which can improve the data transmission speed of AI infrastructure, reduce latency and power consumption, which is crucial for AI infrastructure and applications that require high-speed data transmission.

On June 26th, Eastern Time, at the 2024 Optical Fiber Communication Conference (OFC), Intel's Integrated Photonics Solutions (IPS) group showcased the industry's most advanced and first fully integrated Optical Computing Interconnect (OCI) chip, which is co-packaged and runs real-time data with Intel CPUs.

Intel's OCI chip achieves co-packaged optical input/output (I/O) in emerging AI infrastructure in data centers and high-performance computing (HPC) applications, representing a leap in high-bandwidth interconnects. Intel stated, "We have achieved a revolutionary milestone in integrating optoelectronics into high-speed data transmission."

In terms of functionality, this first OCI chip supports 64 independent channels, each capable of transmitting data at a rate of 32 gigabits per second (Gbps), and efficiently transmits data over a fiber optic cable up to 100 meters, aiming to meet the growing demand of AI infrastructure for higher bandwidth, lower power consumption, and longer transmission distances. It enhances the connection between CPUs and GPUs in clusters, and supports innovative computing architectures such as memory extension and resource decoupling.

Next-generation optical I/O technology is driving innovation in computing platforms to meet the growing AI workloads

With the rapid development of AI technology, applications such as autonomous driving, advanced data analytics, and virtual assistants are becoming increasingly popular globally, leading to a sharp increase in demand for computing resources. In particular, the rapid development of large language models such as GPT and generative AI technologies has greatly promoted the application of AI technology. However, these advanced AI models require massive amounts of data processing and generation, placing high demands on computing resources and data transmission.

As machine learning models continue to scale, their role in AI acceleration becomes increasingly complex, requiring high computational and data processing capabilities to operate effectively. This demand for high-performance computing platforms is driving exponential growth in input/output (I/O) bandwidth and extending data transmission distances.

To address this challenge, data centers are moving towards the use of larger processing unit clusters, such as CPUs, GPUs, and IPUs, as well as more efficient resource utilization architectures, such as xPU decoupling and memory pooling. The implementation of these technologies will improve processing efficiency, reduce system latency, and optimize resource allocation, supporting a wider range of AI computing and applications.

While traditional electronic I/O systems exhibit high bandwidth density and low power consumption when transmitting large amounts of data, their biggest weakness is the short transmission distance, typically limited to within one meter. This severely restricts the layout of devices within data centers, imposing strict spatial constraints on the connection between components To break through this limitation, data centers and early AI clusters have begun to adopt pluggable optical module technology, which can provide longer transmission distances than electronic I/O. However, as AI applications continue to increase in resource demand, the cost and energy consumption pressures of optical modules have also increased.

To address these challenges, a new generation of optical I/O technology has emerged. This technology encapsulates optical I/O with processors (such as CPUs, GPUs, or IPUs, collectively referred to as xPUs), significantly increasing bandwidth, optimizing the internal optical and electrical signal transmission within the chip, significantly reducing energy consumption, and greatly reducing latency in the data transmission process, which is crucial for AI applications that require rapid response.

What's even more exciting is that this technology supports transmission distances far beyond the past, providing greater flexibility in data center design, allowing systems to adapt to a wider range of expansion needs. The promotion of optical I/O technology not only addresses the pain points of data transmission but also paves the way for the future development of AI and machine learning.

To give an analogy, traditional electronic I/O connections are like old-fashioned horse-drawn carriages, efficient in short-distance transmissions but inadequate for long-distance data transmission needs. On the other hand, optical I/O technologies such as Intel's OCI chip are like modern cars and trucks, capable of transmitting more data over longer distances while maintaining data integrity, far surpassing the performance of traditional electronic I/O.

With the continuous expansion of AI and ML model demands, optical I/O, with its outstanding transmission capabilities and efficient energy utilization, has become a key driving force for the future development of AI technology. Just as cars and trucks meet the modern society's needs for fast and large-scale logistics, optical I/O enables data to be transmitted faster and more efficiently over longer distances, which is crucial for expanding AI infrastructure.

Intel Leads in Silicon Photonics

With over 25 years of profound research foundation, Intel Labs has made groundbreaking achievements in integrated photonics. Intel is not only the first enterprise to successfully develop and mass-produce silicon photonics products but also has won the trust of major global cloud service providers with its excellent product reliability.

Intel's core competitiveness lies in its unique hybrid laser wafer technology and direct integration process, which not only improves product reliability but also reduces costs. This unique approach allows Intel to provide excellent performance while maintaining high efficiency. To date, Intel's robust production platform has shipped over 8 million integrated circuit chips, which integrate over 32 million chip-level lasers. The failure rate of its lasers is extremely low (less than 0.1), a widely recognized industry benchmark indicating very low failure rates.

These chips are packaged in pluggable transceiver modules and deployed in large data center networks, serving multiple major cloud service providers for applications at 100Gbps, 200Gbps, and 400Gbps Currently, Intel is still developing the next generation of chips with a speed of 200Gbps per channel to support upcoming applications of 800Gbps and 1.6Tbps.

In terms of manufacturing technology, Intel has introduced a brand new silicon photonics manufacturing process node, which not only improves device performance but also achieves higher integration and better coupling efficiency, while significantly reducing costs. Intel has continuously made breakthroughs in chip laser and SOA performance, cost control, and energy efficiency optimization, reducing chip area by over 40% and energy consumption by over 15%, further consolidating its leading position in the field of silicon photonics technology.

Intel's current OCI chip module is still in the prototype stage. Looking ahead, Intel is collaborating with specific customers to package OCI with their system-on-chips (SoCs) to develop an innovative optical input/output solution