Linkedin Github

Composable Server Infrastructure

Today’s server concepts are still monolithic and are mostly tailored towards one application:

  1. CPU-centric: powerful Multi-socket server CPUs for general workloads and typically lots of memory
  2. Accelerator-centric: Powerful CPU with PCIe attached FPGA or GPU accelerators
  3. Storage-centric: mid-range CPU power with various SSD or HDD storage

 

The allocation of resources is done once at the initial deployment. Run-time changes are not possible, which leads to several disadvantages:

  • Blocked resources: Allocated, but temporary unused resources are not available for other applications
  • Overprovisioning and “Cloud waste”: Appliances allocate more resources than they actually need

Resource Pooling

CAPE’s Edge hardware platforms allow for a fully composable server infrastructure:

  • Fully disaggregated server resources:
    • Compute
    • Storage
    • Acceleration
    • Peripherals
    • NIC
    • Memory
  • Application driven infrastructure
  • Share resources within servers and between servers
  • Flexible communication infrastructure based on PCIe/CXL
CAPE's Edge Server Resource Pools

Compute eXpress Link (CXL) Technology

CXL (Compute Express Link) is the next generation of interconnection standards.

CXL Approach from "Introducing the CXL 4.0 Specification December 4, 2025"

Based on PCIe, CXL offers

  • Cache coherent access between hosts and devices for low latency data sharing
  • Memory expansion and pooling to reduce overprovisioning and improve utilization
  • Seamless attachment of heterogeneous accelerators and peripherals

CXL as a key enabler for Composable Server Infrastructure

  • Boost resource efficiency: Use compute, memory and accelerator resources where they are needed, when they are needed
  • Run-time adaptable server infrastructure to react to changing workloads without hardware changes
  • Interoperability: leverage a growing ecosystem of CXL compliant components and software, backwards compatible to PCIe

CXL in CAPE

  • Both hardware platforms, eHPS and EMDC support CXL
  • CXL configuration is managed by the open hardware management openMPMC
  • The open software stack for deployment supports dynamic resource allocation
  • The Use Cases leverage the coherent memory pools for challenging AI workloads at the Edge

Open Server Management (openMPMC)

The management of heterogeneous Edge Servers comes with multiple challenges. CAPE meets these with openMPMC, providing a single interface for managing multiple servers:

  • Unified management platform
    • Web-based interface for intuitive manual operation
    • Redfish-compatible API for automation and integration
    • Supports eHPS, EMDC, and third-party servers
    • Hardware inventory and sensor monitoring
    • Power control and boot configuration
    • PCIe/CXL resource composition management
  • Bidirectional Redfish integration
    • Acts as Redfish client for BMC communication
    • Serves as Redfish endpoint for external tools
  • Secure remote management
    • Lightweight local agents for isolated networks
    • Outbound gRPC over TLS (no VPN required)
    • Proxy between central platform and BMC endpoints
  • Enterprise features
    • Multi-tenant support with role-based access control
    • Flexible deployment via Docker Compose or Kubernetes
openMPMC server management