We are thrilled to announce that CoHDI has officially been accepted as a Cloud Native Computing Foundation (CNCF) Sandbox project!
This acceptance into the CNCF Sandbox marks an important milestone in CoHDI’s journey. We would like to thank the CNCF Technical Oversight Committee (TOC) for their invaluable guidance and support throughout the process and, of course, our team and community for their essential contributions that were key to making this happen.
Launched in March 2025 as a collaborative effort between Red Hat, FSAS, Fujitsu, IBM Research, and NTT, CoHDI (formerly known as InfraDDS) was founded to cultivate a community-driven, standards-based ecosystem for next-generation architectures built on Composable Hardware in Disaggregated Infrastructure (CoHDI, pronounced “Cody”) for Kubernetes nodes. CoHDI enables host-level dynamic attachment and detachment of PCIe devices through Dynamic Resource Allocation (DRA) on Kubernetes nodes within a composable disaggregated infrastructure. This facilitates flexible and efficient resource allocation across various use cases. For instance, in LLM inference with Prefill/Decode Disaggregation, it accommodates the distinct resource requirements of the prefill phase (compute-bound) and the decode phase (memory-bound). Similarly, in Agentic AI workflows, it dynamically adapts to the varying resource needs of each operational phase. Ultimately, CoHDI holds great potential for driving highly energy-efficient and sustainable operations.
“ As a feature subset of NTT’s IOWN Data-Centric Infrastructure (DCI) for scale-up/down for each compute node, we believe that CoHDI software can bridge Kubernetes-driven cloud-native operations and hardware innovation within composable disaggregated infrastructure of multi-vendor solutions. Open collaboration on sustainable compute hardware system operations is essential to building flexible, future-proof infrastructure in the era of AI. We are proud to support this effort by contributing to the CoHDI ecosystem. ”
– Shintaro Mizuno, VP of Software Innovation Center, NTT
What CoHDI brings to the CNCF landscape
The CoHDI software suite is designed to bridge the gap between traditional hardware boundaries and cloud-native orchestration. By integrating directly with Kubernetes’ Dynamic Resource Allocation (DRA) and collaborating with SIG Node, SIG Autoscaling, and SIG Scheduling, CoHDI seamlessly manages disaggregated resources through three core components:
- Composable-DRA-Driver
A component that works in conjunction with the Dynamic-Device-Scaler to enable dynamic device scaling based on Pod requests. It acts as a bridge between Kubernetes and the CoHDI manager for composable disaggregated infrastructure, exposing available resources managed by the CoHDI manager as ResourceSlices within the Kubernetes DRA framework.
- Dynamic-Device-Scaler
A component that works alongside the Composable-DRA-Driver to enable dynamic device scaling (adding and removing) based on Pod requests, without requiring OS reboots.
- Composable Resource Operator
A Kubernetes operator that dynamically attaches and detaches composable hardware resources (such as GPUs) to cluster nodes by leveraging the external API of the CoHDI manager for composable disaggregated infrastructure.
Get involved: Follow the “well-lit paths”
Welcome to the CNCF, CoHDI! We look forward to building the future of Composable Disaggregated Infrastructure together.
We invite developers, platform engineers, and researchers to join us in shaping the future of open composable disaggregated infrastructure for Kubernetes.
Please feel free to reach out to us on the CNCF Slack, or via Github for any feedback, enhancement requests, or use case discussions!
Facts Only
CoHDI has been accepted as a Cloud Native Computing Foundation (CNCF) Sandbox project.
CoHDI was launched in March 2025.
Founding collaborators include Red Hat, FSAS, Fujitsu, IBM Research, and NTT.
CoHDI was formerly known as InfraDDS.
The software enables dynamic attachment and detachment of PCIe devices on Kubernetes nodes.
The system utilizes Kubernetes Dynamic Resource Allocation (DRA).
CoHDI consists of three core components: Composable-DRA-Driver, Dynamic-Device-Scaler, and Composable Resource Operator.
The project aims to support architectures built on Composable Hardware in Disaggregated Infrastructure.
Application examples include LLM inference (Prefill/Decode Disaggregation) and Agentic AI workflows.
Shintaro Mizuno, VP of Software Innovation Center at NTT, provides a supporting statement.
The project integrates with SIG Node, SIG Autoscaling, and SIG Scheduling.
Executive Summary
CoHDI, a collaborative initiative between Red Hat, FSAS, Fujitsu, IBM Research, and NTT, has transitioned into the CNCF Sandbox. Originally known as InfraDDS, the project focuses on bridging cloud-native orchestration with composable disaggregated hardware. By utilizing Kubernetes Dynamic Resource Allocation (DRA), CoHDI allows for the host-level dynamic scaling of PCIe devices without requiring operating system reboots.
The technical framework relies on a three-part suite consisting of a Composable-DRA-Driver, a Dynamic-Device-Scaler, and a Composable Resource Operator. These tools are designed to optimize resource efficiency for high-demand AI workloads, specifically addressing the differing compute and memory needs of LLM prefill and decode phases. While the project is currently in the sandbox stage—implying it is in an early phase of incubation—it seeks to establish a standards-based ecosystem to improve energy efficiency and sustainability in large-scale compute environments.
Full Take
The strongest version of this narrative is that the industry is hitting a physical wall where static hardware allocation cannot keep pace with the volatile resource demands of generative AI. By decoupling hardware from specific nodes, CoHDI proposes a "liquid" infrastructure that maximizes utilization and reduces waste.
This is a classic vendor-led ecosystem play. The narrative relies heavily on the "Industry Consensus" pattern, where the collaboration of five major corporate entities is used to signal inevitable adoption and stability. The terminology—"well-lit paths," "cloud-native," and "composable"—serves as a professional shorthand that frames this specific technical approach as the only logical evolution for AI infrastructure.
Root Cause: This reflects a shift from "Virtualization of Software" to the "Virtualization of Physical Hardware." The unstated assumption is that hardware disaggregation is the primary bottleneck for AI efficiency, rather than software optimization or algorithmic efficiency.
Implications: If successful, power shifts further toward the providers of the orchestration layer and the specialized hardware managers. While it promises energy sustainability, the second-order effect may be an increase in systemic complexity, creating a higher barrier to entry for smaller players who cannot afford the proprietary "composable" fabrics required to make this software useful.
Bridge Questions:
1. How does the latency introduced by a disaggregated hardware fabric compare to traditional direct-attach PCIe devices in real-world LLM performance?
2. Which specific hardware standards must be adopted across vendors for this "multi-vendor solution" to actually function without proprietary lock-in?
Counterstrike Scan: A coordinated campaign would use "sustainability" and "open standards" as a moral shield to push a specific hardware architecture that favors the founding members' product roadmaps. The current text is a standard corporate announcement and does not exhibit the aggressive urgency or deceptive framing typical of an influence operation.
Patterns detected: ARC-0045 Authority Game
Sentinel — Human
The text reads like a high-level, professionally written announcement combining marketing language with precise, technically dense details, strongly suggesting human authorship informed by deep subject matter expertise.