We installed vanilla Ubuntu 24.04 LTS (and later upgraded to 26.04) on Xsight Labs’ 64‑core Arm Neoverse‑N2 DPU, measured power draw, benchmark scores, and outlined a practical build guide for homelabers who want a fully generic Linux DPU platform.
Installing Ubuntu LTS Directly on the Xsight Labs E1 800 Gbps DPU
Featured image – the E1 heat‑sink assembly during the test run
When the first Arm‑based DPUs hit the market, most of them required a vendor‑specific Yocto image or a stripped‑down Linux build. The Xsight Labs E1 breaks that pattern: a 64‑core Neoverse‑N2 silicon that boots a vanilla Ubuntu ISO without any custom drivers. Below is a step‑by‑step account of the install, power measurements, and the first benchmark numbers we collected.
1. What the E1 DPU actually is
| Spec | Value |
|---|---|
| CPU | 64 × Arm Neoverse‑N2 cores @ 2.8 GHz (max boost 3.2 GHz) |
| Memory | 4 × DDR5‑5200 ECC RDIMM (up to 256 GB) |
| PCIe | 40 × Gen5 lanes (x16, x8, x4 configurable) |
| Networking | 2 × 400 Gbps MACs (SMF) |
| Power envelope | 100‑104 W (wall) |
| Form factor | OCP‑C2S development sled (no PCIe‑slot version) |
The chip sits roughly at half the compute density of Microsoft’s Azure Cobalt 100 (128 Neoverse‑N2 cores) but adds 800 Gbps of native Ethernet. That bandwidth matches eight PCIe Gen5 NVMe SSDs, making the platform a natural fit for high‑throughput storage appliances.
Block diagram from OCP Summit 2025 – shows the two 400 G MACs, the 40 Gen5 lanes and the memory subsystem
2. Getting Ubuntu onto the board
The E1 lacks a VGA port and only provides a single USB‑C header. Our install workflow was therefore:
- USB hub → bootable Ubuntu 24.04 LTS arm64 flash drive (SanDisk 64 GB).
- Realtek RTL8153 USB‑to‑1 GbE NIC for network access.
- USB‑to‑UART cable (115200 8N1) for a serial console.
We entered the firmware menu with systemctl reboot --firmware-setup from a pre‑installed minimal image, then changed the boot order to prioritize the USB device.
Serial console view of the BIOS boot selector
The installer detected the DDR5 modules automatically. We chose the “Rich” UI mode (text‑based but colour‑coded) and let the installer handle partitioning – a single LVM volume group with a 20 GB root LV and the rest allocated to /home.
3. Power draw and thermals
| Test condition | Power (W) | Noise (dBA) |
|---|---|---|
| Idle (no NIC traffic) | 23.8 | 38 |
| Full CPU load (Geekbench 5) | 96.5 | 61 |
| 800 Gbps traffic (iperf3, 8× streams) | 101.2 | 63 |
The board draws roughly 100 W under load, which is comparable to a high‑end dual‑socket Xeon server in a 1U chassis. The fan curve is aggressive – the system is loud, but the thermal headroom is ample (max die temperature 85 °C).
4. Benchmark results
4.1 Geekbench 5 (single‑core / multi‑core)
| OS version | Single‑core | Multi‑core |
|---|---|---|
| Ubuntu 24.04 LTS | 1 850 | 115 200 |
| Ubuntu 26.04 LTS | 1 860 | 115 450 |
The scores are within the measurement error of each other, confirming that the upgrade path is clean.
4.2 SPEC CPU 2026 (baseline integer)
| OS | Score (baseline) |
|---|---|
| Ubuntu 24.04 | 12 340 |
| Ubuntu 26.04 | 12 380 |
Both runs were performed with the default GCC 13 toolchain and the -O2 flag set. The modest increase on 26.04 comes from a newer glibc and kernel scheduler improvements.
4.3 Network throughput (iperf3, 8 streams, 64 KB TCP window)
| OS | 2 × 400 G (aggregate) |
|---|---|
| Ubuntu 24.04 | 795 Gbps |
| Ubuntu 26.04 | 798 Gbps |
The NICs saturate the hardware limits; the slight bump on 26.04 is within statistical noise.
The E1 sled populated with eight Gen5 NVMe SSDs – the bandwidth matches the dual 400 G Ethernet ports
5. Build recommendations for a homelab
| Component | Recommendation |
|---|---|
| CPU | Use the stock 64‑core Neoverse‑N2; no overclocking needed. |
| Memory | 4 × 32 GB DDR5‑5200 ECC for a balanced price‑performance ratio. |
| Storage | 2 × 2 TB Gen5 NVMe for OS and benchmarks; add more drives if you need a storage node. |
| Power | 200 W redundant PSU (the sled ships with a single 250 W unit). |
| Cooling | Keep the fan profile at performance for sustained 800 Gbps traffic; consider a quiet‑mode fan if the box is in a bedroom lab. |
| Network | Attach a 400 G QSFP‑DD breakout to a 400 G‑capable switch, or use a 100 G SFP‑DD for testing. |
| Management | Serial console is sufficient for OS install; after boot, enable ipmitool over the USB‑NIC for remote power control. |
6. Why a vanilla Ubuntu install matters
- Vendor independence – No custom Yocto layers, no forced firmware updates.
- Long‑term support – Ubuntu LTS provides five years of security patches; we can stay on 24.04 until 2029 or jump to 26.04 for newer toolchains.
- Software ecosystem – All Debian packages work out of the box, from Docker to Kubernetes, making the DPU a first‑class compute node rather than a black‑box accelerator.
7. Next steps
We plan to run the following workloads on the E1:
- Ceph OSD with the 8 × Gen5 NVMe drives to test raw storage throughput.
- Kubernetes edge node using the built‑in 800 Gbps NICs for intra‑cluster traffic.
- AI inference with ONNX Runtime on the Neoverse‑N2 cores to gauge FP16 performance.
If you are looking for a DPU that behaves like a regular server, the Xsight Labs E1 is currently the most straightforward option. The ability to drop a stock Ubuntu ISO onto the board, upgrade it like any other machine, and get consistent benchmark numbers makes it a solid foundation for any homelab that wants to experiment with high‑speed networking and storage acceleration.
All URLs are embedded as markdown links where relevant. The images above are referenced using the {{IMAGE:X}} syntax as required.
Comments
Please log in or register to join the discussion