²ÝÁñÉçÇø

How Are Data Centers Evolving?

Cloud and edge computing have grown more pervasive in recent years, with AI quickly becoming the biggest driver of bandwidth. In response to these trends, data centers have transformed from basic servers dealing with a manageable amount of data to complex, sophisticated multi-rack systems that handle far more data than we ever imagined. projects that by 2025, the world will have as much as 175 zettabytes of data, split between the cloud and data centers.

Going forward, data centers are moving towards a disaggregated architecture, where homogeneous resources such as storage, compute, and networking reside in separate boxes connected via electro/optical interconnects. Network architectures themselves are flattening, delivering speed and scalability while driving demand for higher bandwidth and efficient connectivity across longer distances.

Ethernet provides the primary interface for  box-to-box connections. It offers advantages such as speed negotiation and support for different kinds and classes of media (such as optical fiber, copper cables, and PCB backplane). With 1.6TbE, the initial applications will likely be compute clusters within data centers that are aimed at processing unstructured workloads on large language models (LLMs). LLMs burst onto the scene with the likes of ChatGPT. These models have trillions of parameters, and their number of parameters doubles every few months. Much of the data is kept in memory and must be processed together, so the system requires many processors connected via a low-latency network. To process such a workload, entire clusters would act as single compute devices, with multiple clusters together processing terabytes of data. Ethernet offers an ideal connectivity protocol for these massive systems (Figure 1).

²ÝÁñÉçÇø offers a complete solution for 1.6T Ethernet, adopted by multiple customers, designed from the ground up and leveraging our experience and industry leadership developing IP for 400G and 800G. The solution has demonstrated the ability to reduce interconnect power consumption by up to 50% compared to existing implementations. With our expertise in optimizing subsystems, our complete solution enables designers to reduce turnaround time as well as power and latency. The ²ÝÁñÉçÇø 224G Ethernet PHY IP offers excellent signal integrity and jitter performance , with zero post-FEC BER, and additional margin for channel loss. It supports pulse-amplitude modulation 4-level (PAM-4) and non-return-to-zero (NRZ) signaling to deliver up to 1.6T Ethernet. The new multi-rate Ethernet MAC and PCS controllers with patented Reed Solomon forward error correction (FEC) architecture cut latency by 40% compared to prior generations. Knowing that the configurable Ethernet PHY and controller IP are tested and interoperable, engineers can focus their time on differentiating their designs. Figure 3 shows  interoperability demonstrations of the 224G Ethernet PHY IP at various industry events.  

²ÝÁñÉçÇø also provides the industry¡¯s first 1.6T Verification IP (VIP) enabling early RTL verification, SoC bring-up, and system-level validation, offering designers a fast path to design verification closure. Leveraging our technical expertise, participation with standard committees, and collaboration with key ecosystem and silicon partners, we continue to expand our portfolio, covering automotive BaseT, AVB/TSN, USXGMII, FlexE, and all speed modes up to 1.6T Ethernet. Figure 4 highlights our 1.6T Ethernet IP solution.