As AI clusters grow larger, this connectivity approach aims to overcome bottlenecks limiting performance, scalability, and efficiency.
Microchip Technology has introduced its XpressConnect PCIe 6.0 and CXL 3.1 retimers, targeting one of the most persistent challenges in modern AI systems: efficient data movement across increasingly complex compute and memory fabrics.
As AI workloads scale, data center designers are facing growing limitations in signal reach and rising latency that can leave high value memory and compute resources underused, particularly across large GPU clusters. These issues become more pronounced as interconnect speeds increase to 64 GT per second, where signal integrity constraints can restrict system scale and increase architectural complexity.
The retimers are designed to extend electrical signal reach beyond conventional PCIe Gen 5 and Gen 6 limits, enabling more flexible system configurations across baseboards, riser cards, and cabled interconnects. They are also engineered to operate within tight power and thermal budgets required in dense AI infrastructure deployments.
A key feature of the XpressConnect devices is very low pin to pin latency, rated at under 12 nanoseconds, which Microchip states is around 80 percent lower than PCIe 6.0 specification expectations. This reduction in latency is intended to improve utilization of GPUs and AI accelerators by reducing data stalls across high density compute environments.
The retimers integrate with Microchip’s broader ecosystem including PCIe Gen 6 switches, storage controllers, and NVMe solutions, helping create a more unified and interoperable data center fabric. They also support ChipLink diagnostic tools for real time monitoring and simplified system troubleshooting, improving reliability and reducing operational complexity in AI deployments.
“AI data centers are increasingly constrained not by compute, but by the ability to move data efficiently across the system. As PCIe 6.0 pushes speeds to 64 GT per second, signal reach and latency become critical design challenges,” says Brian McCarson, corporate vice president and general manager of Microchip’s data center solutions business unit. “Our XpressConnect retimers are designed to act as the high performance nerve center of the AI server, helping customers build more scalable and power efficient fabrics by reducing latency and improving connectivity across dense GPU clusters. This system level approach allows data center architects to reclaim underused resources and improve overall platform efficiency at scale.”
Click here for the official announcement.
