Micas Networks

CPO: Leveraging Silicon Reliability for AI Networking At Micas, we are committed to revolutionizing AI Networking through innovative, reliable, and cost-effective optical solutions. CPO is emerging as a transformative technology that integrates optical components directly with switch or compute silicon, thereby harnessing the inherent reliability of silicon manufacturing. This integration not only reduces power consumption and link latency but also provides a more robust foundation for next-generation data centers. While PAM4 technology has been critical in increasing data transmission rates, it comes with several limitations. For example, as data centers demand higher performance—reaching 400G and 800G speeds—the power consumption of traditional PAM4-based pluggable optical modules rises dramatically. This increased power draw places significant pressure on overall energy efficiency and operational costs. Moreover, PAM4 solutions often face challenges with signal integrity; signal loss becomes an issue over long distances and through multiple copper channels, particularly as SerDes transmission rates increase. Additionally, electrical channel design constraints can hinder the performance of Linear Drive Pluggable Optics (LPO) solutions, making it difficult to meet the upcoming requirements for 224G SerDes. In contrast, CPO offers compelling advantages that directly address these limitations. By integrating optical and electrical components within a single package, CPO reduces the length of the electrical path, which not only enhances signal integrity but also significantly lowers latency. This streamlined design results in fewer potential points of failure and enables data centers to operate more reliably with minimal downtime. Data from industry leaders indicates that CPO can deliver substantial power savings compared to DSP pluggable and Linear Receive Optics (LRO) solutions, making it a more sustainable option for high-density AI networking. Furthermore, CPO improves scalability and networking density, paving the way for applications beyond 100T. The ability to support larger scale architectures and deliver greater performance per bit is critical as data centers evolve to meet ever-growing data demands. In addition to performance and reliability benefits, the integration provided by CPO also drives significant cost reductions—often achieving cost savings of more than 30% compared to conventional pluggable optics. In summary, while PAM4 modulation has enabled higher data rates, its inherent limitations in power consumption, signal degradation, and design constraints can hinder future growth. CPO overcomes these challenges by leveraging silicon’s superior reliability, ensuring enhanced performance, and reducing overall system costs. At Micas, we are proud to lead the way in delivering next-generation AI networking solutions that set a new standard for efficiency and reliability in modern data centers.

CPO: Leverage Silicon Reliability, driving innovation one decision point at a time. In our recent Llama 3 405B model training failure report, network issues (link flaps is #1 cause) accounted for 8.4% of the overall failures. This finding highlights a persistent challenge: intermittent network connections that disrupt data flow and impair system performance. But what exactly are link flaps? They occur when a network connection repeatedly goes up and down, leading to erratic data transmission. Traditional DSP pluggable transceiver solutions require three distinct decision points at the PCB, within the ASIC, and at the cage. At each of these junctures, the slicer must determine where to sample the incoming signal, and every sampling decision introduces a probability of error. Over time, these compounded errors lead to frequent link flaps; a problem that has only worsened over the years. Enter Co-Packaged Optics (CPO). By integrating the optical components directly with the switch ASIC, CPO consolidates the decision-making process to a single point at the receiver’s SerDes. This streamlined approach reduces the potential error opportunities by roughly 67%, dramatically lowering the incidence of link flaps. Why is this reduction so crucial? Consider the benefits: * Increased Reliability: Fewer link flaps mean a more stable, dependable network that ensures uninterrupted data flow. * Improved Performance: Stable connections translate into consistent data throughput and reduced latency, which are critical for both real-time applications and large-scale data processing. * Reduced Operational Overhead: With fewer intermittent failures, network teams spend significantly less time troubleshooting and maintaining infrastructure, thereby lowering operational costs. Moreover, the simplified optical interface not only boosts reliability but also paves the way for more scalable and cost-effective network designs. When we link this back to our training failure report; where 8.4% of issues were network-related; it’s clear that addressing link flaps isn’t just about network performance; it has direct implications for the success of high-performance computing and AI training. In today’s competitive market, where every percentage point matters, CPO’s ability to significantly reduce error occurrences offers a compelling edge; delivering enhanced network stability, higher efficiency, and lower costs.

CPO: Stop Burning Cash—Unlocking more power for compute: Co-Packaged Optics integrates the optical engine directly with the switch chip in a single package. By eliminating traditional pluggable optics and shortening electrical signal paths, CPO minimizes power dissipation and enhances overall efficiency. This integration not only reduces energy consumption but also simplifies thermal management and improves signal integrity. Here are some key power-saving benefits of CPO: 1. 50% Reduction: CPO is expected to cut the power consumption of existing pluggable optical module architectures by up to 50%. Micas CPO Switch using Broadcom’s TH5-Bailly delivers an impressive energy consumption of less than 6 pJ/bit. 2. No DSPs: Traditional pluggable solutions often rely on digital signal processors (DSPs) to restore signal integrity, contributing significantly to overall power usage. CPO bypasses the need for these DSPs, reducing both complexity and power draw. 3. 80% Saving: By integrating optical and electrical components, CPO achieves more than 80% savings in insertion loss compared to conventional pluggable modules, further enhancing efficiency. According to Broadcom’s analysis in a 30,528 GPU cluster, CPO can save an estimated 466 kW compared to DSP-based pluggables. As the industry moves towards 100T+ applications, CPO is poised to play a crucial role in enabling high-density, energy-efficient optical interconnects, revolutionizing data center infrastructure and driving a more sustainable future. Discover Micas' cutting-edge optical networking solutions at OFC 2025. Our partnership with Broadcom, currently supplying CPO switches to the US hyperscale data center, demonstrates our commitment to delivering unmatched performance and power savings for your network. #CoPackagedOptics #CPO #DataCenter #PowerSavings #OpticalInterconnects #Broadcom #AI #Networking #Innovation Source: Broadcom documentation and analysis via industry reports.

Why Co-Packaged Optics (CPO) is the Future of AI Infrastructure: Boosting GPU Utilization and Energy Efficiency The rapid advancement of LLM is reshaping the technological landscape, driving unprecedented demands for faster, more efficient, and scalable infrastructure. Boosting GPU Utilization with CPO （A Must-Have Beyond 200G Per Channel） For the past 10 years, we've been striving to improve the utilization of what is now the most expensive computing device: the GPU. However, their performance is often bottlenecked by the speed at which data can be transferred between GPUs, CPUs, and memory. CPO addresses this challenge by enabling faster data transfer and lower latency. By placing the optical engine closer to the switch chip, CPO reduces the distance data must travel, allowing GPUs to communicate more efficiently. This results in higher GPU utilization, as less time is spent waiting for data and more time is spent processing it. For AI workloads, this translates to faster training times, more accurate models, and the ability to tackle even larger datasets. Energy Efficiency: A Game-Changer for Data Centers (Reduce power by more than 50%) Energy efficiency is a critical concern for data centers, especially as AI workloads continue to grow in scale and complexity. Traditional pluggable optics consume significant amounts of power, contributing to rising operational costs and environmental impact. CPO offers a compelling solution, with the potential to reduce power consumption by more than 50% compared to traditional modules. By integrating the optical engine and switch chip, CPO minimizes the need for power-hungry electrical-to-optical conversions. This not only reduces energy usage but also lowers heat generation, simplifying cooling requirements and further reducing costs. For data centers supporting AI workloads, CPO represents a win-win: improved performance and significant energy savings. Conclusion The rise of AI is driving a paradigm shift in how we think about computing infrastructure. Co-Packaged Optics (CPO) is at the forefront of this shift, offering a powerful solution to the challenges of GPU utilization, energy efficiency, and scalability. By integrating optics and electronics into a single package, CPO is unlocking new levels of performance and efficiency, paving the way for the future of AI. The adoption of CPO is already gaining momentum, with major tech companies like TSMC, Broadcom and Nvidia investing heavily in its development. As we look ahead, it’s clear that CPO will be a cornerstone of AI infrastructure, enabling faster, more efficient, and more sustainable computing. For those in the AI field, understanding and embracing this technology will be key to staying ahead in an increasingly competitive landscape.

The Evolution of CPO (Co-Optics): Redefining GPU to GPU communication As the VP of Strategy at Micas, I’ve witnessed the remarkable journey of CPO—from early research breakthroughs to today’s high-capacity data centers. Now, we stand on the brink of another pivotal transition: “co-optics,” a paradigm where optics and electronics are merged within the same package to overcome performance bottlenecks imposed by traditional copper interconnects. By reducing signal loss, latency, and power consumption, co-optics is rapidly becoming the must-have technology for high-performance computing and next-generation networks. 1. The 1980s: The Conceptual Birth of Silicon Photonics Laid the Foundation of Integrating photonic with standard Semi Processes 2. The 1990s: Waveguide Innovations and Prototype Systems Waveguides are critical for routing light signals. Overcoming early challenges—such as high propagation losses—validated that photons could be guided effectively on silicon substrates. 3. The 2000s: Commercialization and Data Center Adoption Cisco integrated optical solutions into networking gear, while Broadcom acquired and developed optical transceiver technologies. During this time, TSMC refined its semiconductor fabrication processes, paving the way for large-scale silicon photonics manufacturing. 4. The 2010s: High-Performance Computing and Advanced Integration HPC and AI platforms adopted silicon photonics for node-to-node connections. NVIDIA, began exploring photonic interconnects to optimize data flow in AI clusters. Meanwhile, Intel unveiled transceivers that integrated lasers and modulators on a single chip 5. The Present: Co-Optics Leading the Way By placing lasers, modulators, and processors in one package, the distance between electrical and optical interfaces shrinks dramatically, boosting throughput and energy efficiency. TSMC’s advanced packaging co-packaged designs are becoming a reality thanks to sophisticated foundry technologies like chip-on-wafer-on-substrate (CoWoS) and fan-out packaging. Looking Ahead: From Incremental to Transformational At Micas, we partner with industry leaders Broadcom to develop next-generation technology that exceeds today’s performance and energy requirements. Co-optics addresses fundamental bottlenecks by merging optical and electronic components into a single cohesive unit. It’s particularly critical for modern applications like large-scale AI training and inferencing all of which demand low latency and high bandwidth. Expect further advancements in waveguide design, on-chip laser integration, and thermal management as the industry races to meet these needs. As these technologies mature, we’ll continue sharing insights and illustrating how co-optics solutions make its way into AI data centers. The future is bright for CPO, we’re just getting started.

Exciting News! 🚀🚀🚀 Micas Networks is a proud sponsor of UC San Diego's team competing in the 2024 Student Cluster Competition at the Supercomputing Conference in Atlanta, Georgia, from November 17-22.   We also lent them a high-speed 400G Tomahawk-4 based switch to keep their data moving as fast as possible.   Best of luck to this wonderful group of up-and-coming HPC technologists! Also, big thanks to our fellow sponsors.   #SC24 #Supercomputing #StudentClusterCompetition #Innovation #Technology #TeamSpirit #Tomahawk

🔥 Exciting developments in the world of open networking! In our latest VMblog Q&A, we sit down with Mitch Galbraith, COO of Micas Networks, to discuss how they're disrupting the industry with cutting-edge 51.2T Co-Packaged Optic switches, strategic partnerships, and rapid growth. Micas is setting the pace for the future of #networking, catering to #hyperscalers, cloud data centers, and OEMs with innovative solutions. Curious to learn more about how they’re driving the next wave of disaggregated, AI-enabling networks? Check out the full interview here 👉 https://lnkd.in/gT-7fmSt #Cloud #OCP #TechInnovation #OpenNetworking

The team at ServeTheHome - a trusted resource for reviews, news and analysis of the newest server, storage and networking trends – took a deep dive of our Tomahawk 5 CPO Switch at last week’s OCP Global Summit. ServeTheHome loved how our CPO switch was running traffic on the show floor and captured several in-process photos of the demo. Check out their article here: https://bit.ly/4e4TMYD #OCPGlobalSummit #OCPSummit2024 #OpenNetworking #Tomahawk5 #CPOSwitch

We’re thrilled to officially announce our new 800G Tomahawk 5 Ethernet Switch! Now shipping to customers worldwide, this switch brings ultra-high performance to AI & ML applications. Check out our announcement for all the details: https://bit.ly/3AddfZf #Tomahawk5 #EthernetSwitch #AIApplications #OpenNetworking