Intel has developed a new Optical Compute Interconnect (OCI) chiplet that aims to address the increasing data transmission challenges brought on by generative AI and large language models like GPT-4. This chiplet is part of Intel’s Integrated Photonics Solutions Group and is designed to enhance the scalability of AI infrastructure by leveraging silicon photonics technology to deliver fast and energy-efficient data transmission to computational points such as GPUs and electronic CPUs.
With the rise of generative AI, the demand for data has surged, leading to bandwidth bottlenecks and rising energy costs in data centers. In response, startups like Ayar Laboratories, Celestial AI, and Light matter, along with established companies like Intel, are working to develop solutions to meet these challenges.
Intel’s OCI chiplet is a prototype silicon photonic product that promises high-speed data transmission capabilities. It combines a photonic integrated circuit (PIC) that includes waveguides, amplifiers, and lasers with an electrical integrated circuit (EIC) that features control circuits and drivers. This chiplet aims to minimize the “I/O tax” by handling optical I/O subsystem and conversion tasks at the edge of computation, allowing GPUs and CPUs to focus on heavy computational tasks.
At a recent demo, Intel showcased the chiplet’s data communication capabilities, achieving 4 terabits per second connectivity while consuming only 10W of power. The chiplet’s energy efficiency and scalability potential make it a promising solution for future data transmission needs in AI infrastructure.
One of Intel’s key advantages in this market is its expertise in integrating reliable lasers directly onto silicon chips. This approach enhances energy efficiency and reduces coupling loss compared to external lasers connected via fiber connections. By integrating lasers at wafer scale, Intel aims to build a complete optical subsystem that offers high performance and cost efficiency.
Intel is currently working with select customers to test the integration of the OCI chiplet with their systems on a chip (SoCs) as an optical I/O solution. The company is exploring different implementation paths to optimize computing resources and enhance bandwidth density and performance parameters to meet future demands.
Intel’s experience in silicon photonics R&D and supplying pluggable transceiver modules for data center networking has positioned the company well for success in delivering innovative solutions for AI infrastructure. Intel sees optics playing a crucial role in addressing the growing demands of AI applications and is committed to contributing to the advancement of the field.
Article Source
http://www.optica-opn.org/home/industry/2024/june/intel_showcases_optical-interconnect_chiplet/