PCIe Gen5 + CXL Data Flow: Visualizing the Future of System Architecture

PCIe Gen5 + CXL Data Flow: Visualizing the Future of System Architecture

🔷Understanding how data and memory move through modern platforms is key to building scalable

AI, cloud, and high-performance systems.

•This Sankey diagram illustrates the complete journey of 430 Tbps of traffic across compute, interconnect, memory, and application layers showing

why PCIe and CXL must be designed together.

🔷The Flow Architecture:

Starting with CPUs, GPUs, DPUs, FPGAs, and NVMe SSDs generating 430 Tbps of data, traffic

enters the protocol layer where a critical split occurs.

• PCIe Gen5 carries 250 Tbps (58%) of traditional I/O traffic

• CXL 2.0/3.0 handles 180 Tbps (42%) of cache-coherent memory traffic

This split defines the evolution from device-centric I/O to memory-centric system design.

🔷Interconnect Fabric Enables Scale:

PCIe Gen5 switches, CXL switches, and direct-attach links form the backbone of the platform.

• PCIe focuses on high-throughput DMA, MMIO, and peer-to-peer transfers

• CXL introduces memory pooling and coherency across devices enabling disaggregated and composable systems.

🔷Memory and Processing in Action:

• PCIe paths serve storage, networking, and accelerator data movement.

• In parallel, CXL.mem and CXL.cache enable 90 Tbps of memory expansion and 60 Tbps of cache-coherent access, reducing DRAM pressure while maintaining low latency for compute-intensive workloads.

🔷From Interconnect to Applications:

• AI/ML training consumes 150 Tbps by combining PCIe-attached GPUs with CXL-based memory expansion.

• High-performance storage, real-time analytics, and cloud services are dynamically served across both paths, depending on latency and coherency requirements.

🔷The Key Insight:

• PCIe moves data fast.

• CXL changes how memory is shared.

With 42% of system traffic leveraging CXL, modern platforms shift from fixed,socket-bound memory to flexible, pooled memory architectures.

This transformation not just higher bandwidth is what defines next generation AI and cloud systems.

By Krishnan Babu

Technology Specialist

Embedded