As AI infrastructure continues to scale, moving data efficiently between processors, accelerators, memory, and storage has become a growing challenge for data center architects. To address signal integrity and latency limitations associated with next-generation interconnects, new PCIe® 6.0 and CXL® 3.1 retimers have been introduced to support memory expansion and resource disaggregation in large-scale AI systems.
The retimers are designed for AI servers and high-performance computing platforms operating at PCIe Gen 6 speeds of 64 GT/s. At these data rates, signal degradation across complex board layouts, riser cards, and cabled connections can limit system scalability and complicate server design. The new devices are intended to extend signal reach beyond conventional PCIe Gen 5 and Gen 6 electrical limits while maintaining the bandwidth required for modern AI workloads.
A key focus of the design is reducing latency within large GPU clusters. The retimers deliver less than 12 ns of pin-to-pin latency, approximately 80% lower than PCIe 6.0 specification limits. Lower latency can help reduce data-transfer bottlenecks between compute and memory resources, improving utilization of GPUs and AI accelerators in large-scale deployments.
As AI data centers increasingly rely on memory pooling and resource disaggregation architectures, maintaining high-speed connectivity between distributed resources becomes critical. By extending signal reach while minimizing latency, the retimers are intended to support larger, more flexible AI fabrics without sacrificing performance.
The devices are designed to integrate with existing PCIe infrastructure and can be deployed in PCIe Gen 3, Gen 4, Gen 5, and Gen 6 systems. This backward compatibility can simplify upgrades and reduce development complexity for system designers transitioning to next-generation interconnect standards.
The retimers are also designed to work alongside PCIe Gen 6 switches, RAID controllers, host bus adapters, and NVMe controllers as part of a broader data center connectivity ecosystem. The goal is to provide an interoperable platform that reduces integration effort while supporting high-bandwidth compute and storage architectures.
For monitoring and troubleshooting, the retimers connect to a diagnostic environment that provides real-time visibility into link performance. Features include graphical eye-diagram capture and four-level pulse amplitude modulation (PAM4) telemetry, allowing engineers and operators to evaluate signal quality and identify potential issues before they affect system performance.
The devices support multiple link bifurcation configurations, including 1×16, 2×8, and 4×4 topologies, enabling designers to tailor connectivity to specific server architectures. Additional enterprise-focused features include hot-plug support and end-to-end data integrity protection.
As AI clusters continue growing in size and complexity, technologies that extend signal reach while minimizing latency are becoming increasingly important. Retimers have emerged as a critical building block for maintaining high-bandwidth connectivity across dense AI fabrics, helping data center operators improve resource utilization and system efficiency while supporting next-generation PCIe and CXL architectures.
Learn More:
