PacketFive | Cnuas Virtual AI/HPC Infrastructure Cnuas Virtual AI HPC Infrastructure Datasheet Full platform reference, switches, NICs, GPUs, fabric and host requirements Document DS-CNU-001 Revision D Issued 12 August 2026 Status Published ============================================================================== Item Value --------- ----------------------------------------------------- Part Cnuas Virtual AI HPC Infrastructure Type Software-emulated virtual AI and HPC infrastructure Version Platform 2026.07 (9fbdddf-dirty) Status Active development Product cnuas.io 1. Overview Cnuas (Irish Gaeilge for "cluster") is the Cnuas Virtual AI HPC Infrastructure, a 100% software-emulated virtual AI and HPC infrastructure for teaching and developing HPC, AI, and RDMA networking (RoCEv2 and native InfiniBand) without physical GPU clusters or InfiniBand switches. Every device is a QEMU/userspace model presented to unmodified guest software, so a virtual GPU or NIC behaves like the real part it emulates. Cnuas is developed under the internal Cnuas engineering programme. It composes four emulated device families, CnuasSwitch (TOR fabric switch), CnuasNIC (RDMA NIC), CnuasGPU (SIMT accelerator), and CnuasLink (GPU fabric), over a forked QEMU and a forked Linux kernel, orchestrated as VMs on a single Ubuntu management host. Components usable without a guest The full rack runs as virtual machines, but three of the four device families are also usable on their own, as host processes with no hypervisor, no guest and no kernel module. This matters when only one part is wanted, for example an accelerator inside an application, or the fabric alone for protocol work. Component Usable with no guest Form ------------- ---------------------- ----------------------------------------------------------------------------------------------------------------- CnuasGPU Yes Soft-GPU in the calling process, including a ggml backend for ggml based applications CnuasSwitch Yes cnuas-vswitchd host daemon; ports are UNIX sockets any host process can attach to CnuasLink Yes cnuasgpu-link-switchd host daemon CnuasNIC No Needs a guest, because the drivers bind an emulated PCI function and the provider uses the kernel ib_uverbs ABI Key features - Full RDMA fabric, RoCEv2 and native InfiniBand, end to end. - SIMT virtual GPUs with a CUDA-style runtime and multi-architecture SIMD backends. - Separate GPU-to-GPU fabric (CnuasLink) modelling an NVSwitch/NVLink-class interconnect. - In-switch Subnet Manager, the fabric comes up without external opensm. - Runs on a single commodity x86-64 host; no special hardware required. - Consistent management plane spanning CLIs, JSON control sockets, and a browser dashboard. - Multi-rack topology modelling on the Open Compute Project Open Rack v3 reference. 2. Rack-level block diagram Two-rack reference deployment. Each rack has a CnuasSwitch top-of-rack (TOR) for the RoCE/IB fabric and a rack-local CnuasLink switch for the GPU fabric. VMs attach to both fabrics; racks are joined by an inter-TOR link. [Diagram omitted from the plain text edition. See the PDF or the online documentation.] 2.1 Rack elevation (ORv3, 24 OU) Logical OU (OpenU) slot assignment for a single rack. CnuasSwitch (TOR) and the rack-local CnuasLink switch occupy the top OUs; compute VMs fill the remainder. Two such racks form the reference deployment above. [Diagram omitted from the plain text edition. See the PDF or the online documentation.] 3. Host / node block diagram Inside a single physical host, the switch daemons run in host userspace; each VM is a QEMU process exposing a cnuas-vnic (NIC) and a cnuasgpu (GPU) PCIe device. Devices connect to the daemons over UNIX-domain sockets (UDS). [Diagram omitted from the plain text edition. See the PDF or the online documentation.] 4. Guest software stack Per-VM stack, from application down to the QEMU device backends. The RDMA path uses stock libibverbs, which dlopen()s the CnuasNIC provider; the GPU path uses libcnuasrt over the cnuasgpu.ko char device. [Diagram omitted from the plain text edition. See the PDF or the online documentation.] 5. Platform specifications Parameter Value ------------------------- ----------------------------------------------------------------------------- Emulation model QEMU custom PCIe devices + userspace switch daemons Host OS Ubuntu 24.04 LTS (x86-64) Guest kernel Linux fork cnuas-v6.19.0 (6.19.0-cnuas+) Hypervisor QEMU fork cnuas-qemu-v0.1.0 Inter-VM data transport UNIX domain sockets (SOCK_SEQPACKET); ivshmem shared memory (GPU fast path) Management transport JSON over UNIX/TCP sockets Reference topology Open Compute Project Open Rack v3 (ORv3), up to 24 OU Fabric switch ports 10 per CnuasSwitch (8 fabric + uplink + console) GPU fabric ports 8 per CnuasLink (configurable, max 32) Build system Bazel (daemons/CLI/WebUI/packaging), Make (kernel, QEMU) Languages C (GNU C, daemons + kernel), Python 3.11+ (tooling), C++ (QEMU models) Versioning Semantic versioning per component (cnuas-vMAJOR.MINOR.PATCH) 6. Component summary Component Part / daemon Kernel Userspace Datasheet --------------- ----------------------- ------------------------------ ----------------------------------------- ------------- Fabric switch cnuas-vswitchd , (host daemon) cnuas-cli, WebUI CnuasSwitch RDMA NIC cnuas-vnic cnuas_net.ko, cnuas_ib.ko libcnuas-rdmav34.so CnuasNIC GPU cnuasgpu cnuasgpu.ko (/dev/cnuasgpuN) libcnuasrt.so, libcnuasdev.so, cnuassmi CnuasGPU GPU fabric cnuasgpu-link-switchd , (host daemon) cnuaslink-cli CnuasLink 6.1 Component versions Live from each submodule's git tag at documentation build time: @@VERSION_TABLE@@ 7. Management and control plane Interface Description ---------------------- --------------------------------------------------------------- cnuas-cli Switch/fabric management CLI (port status, SM, DCB) cnuaslink-cli CnuasLink GPU fabric management CLI cnuassmi / cnuas-smi GPU inventory and telemetry (nvidia-smi style) cnuas-tools VM lifecycle: image/package build, vm up/down/ssh/console/lab cnuas-webui Browser dashboard (FastAPI backend + web frontend) Control sockets JSON over UNIX socket (/var/run/cnuas/, /var/run/cnuaslink/) 8. Host requirements Requirement Minimum Recommended ---------------- ------------------------------------------------ -------------------------------- CPU x86-64 with AVX2 AVX-512 (for CnuasGPU compute) Memory 16 GB 64 GB+ (scales with VM count) Virtualization KVM enabled KVM + nested virt OS Ubuntu 24.04 LTS Ubuntu 24.04 LTS Toolchain GCC (GNU C), Make, Bazel, Python 3.11+, pytest + mkdocs-material for docs 9. Roadmap summary Feature availability at a glance. Full Epic/Task breakdown, status, and the GitHub Projects board are in the Product Roadmap. Workstream Available today Planned ------------------------ ---------------------------------------------------------------------------------------- -------------------------------------------------------------- CnuasNIC (RDMA NIC) RoCEv2 + native IB, RC/UD/SMI/GSI, atomics, SRQ, multicast, ICRC, in-switch SM interop SA path-record, PMA/perfquery, opensm compat CnuasSwitch (fabric) 10-port RoCE/IB, DCB, FDB/LFT, in-switch Subnet Manager, telemetry Switch-side SA/PMA alignment CnuasGPU (accelerator) SIMT runtime, libcnuasrt/libcnuasdev, AVX2/AVX-512 SGEMM, multi-GPU Backend dlopen split, device-backend refactor, CnuasIR, FPGA CnuasLink (GPU fabric) Frame protocol, FDB, discovery, sysfs link state, CLI Collective offload, bandwidth modelling Platform / tooling VM lifecycle CLI, packaging, submodule split, CI Documentation pass 10. Ordering / integration information Item Value ------------------- ------------------------------------------------------------------------- Superproject repo PacketFive/cnuas (git submodule superproject) Submodules linux, qemu, src/cnuasnic, src/cnuasswitch, src/cnuasgpu, src/cnuaslink License Apache-2.0 (see repository LICENSE) Quick start ./build_and_test.sh Documentation https://cnuas.io/docs/ Project site https://cnuas.io 11. Revision history Revision Date Notes ---------- ------------ ---------------------------------------------------------------------- A 2026-07-05 Initial platform datasheet B 2026-07-05 Added rack/host/stack diagrams, detailed specs, component driver map C 2026-07-05 Added rack elevation, live version table, roadmap summary D 2026-08-12 Recorded which components are usable without a guest ============================================================================== PacketFive, Packet Five Networks Ltd., Dublin, Ireland. Cnuas Virtual AI/HPC Infrastructure is published at https://github.com/PacketFive/cnuas under the Apache License 2.0; read it at https://github.com/PacketFive/cnuas/blob/main/LICENSE. Cnuas Virtual AI/HPC Infrastructure is emulation software. It is not affiliated with, endorsed by, or derived from any hardware vendor, and every device it models is a software artefact. The work is published by its authors in a personal capacity and is not sponsored or endorsed by any employer. Specifications describe the referenced revision of the software and may change without notice. Contact info@packetfive.com.