I’ve had my Spark less than a week and any time I put a moderate load on it it hangs and stops responding until I cycle the power. I’ve had this happen with multiple applications. Within a few minutes of being under load it dies.
Component
Value
Device
NVIDIA DGX Spark (GB10 SoC)
DGX OS Version
7.5.0 (OTA updated June 26, 2026)
Kernel
6.17.0-1021-nvidia
GPU
NVIDIA GB10 (Device ID: 2e12, rev a1)
Power Delivery Warnings
[ 1.599851] mlx5_core 0000:01:00.0: mlx5_pcie_event:326:(pid 12): Detected insufficient power on the PCIe slot (27W).
[ 2.173313] mlx5_core 0000:01:00.1: mlx5_pcie_event:326:(pid 366): Detected insufficient power on the PCIe slot (27W).
[ 2.758723] mlx5_core 0002:01:00.0: mlx5_pcie_event:326:(pid 12): Detected insufficient power on the PCIe slot (27W).
[ 3.315569] mlx5_core 0002:01:00.1: mlx5_pcie_event:326:(pid 366): Detected insufficient power on the PCIe slot (27W).
I’ve confirmed all the firmware software is up to date. I did open a support ticket, but I figured I’d post here as well.
The power delivery warnings are benign and do not reflect any performance errors. What kind of load are you running? After rebooting, please run journalctl -k -b -1 -e to collect previous boot logs and please share with me.
Thank you, I forgot to ask for an nvidia-bug-report as well so please share that.
You appear to be encountering an Out-Of-Memory error, which is hanging your system. You should update your Spark to the latest version, and have at least GPU driver 580.159.03. This should stop the device hanging when an application uses too much memory, and will instead kill the process.
I would recommend you change your workload to not limit your memory usage as well, so that other functions have enough memory to be usable.
ii cuda-nsight-compute-13-0 13.0.3-1 arm64 NVIDIA Nsight Compute
ii cuda-nsight-systems-13-0 13.0.3-1 arm64 NVIDIA Nsight Systems
ii cuda-nvtx-13-0 13.0.85-1 arm64 NVIDIA Tools Extension
ii dgx-dashboard 0.29.0 arm64 NVIDIA DGX Dashboard
ii dgx-oobe 0.25.1 arm64 NVIDIA DGX OOBE Core Services
ii dgx-spark-oobe-customize 0.17.9 arm64 NVIDIA DGX Spark OOBE Customizations
ii doca-host 3.2.1-044000-25.10-ubuntu2404 arm64 Software package including drivers, libraries and tools installed on the host server to support NVIDIA networking platforms
ii hpc-sdk-repo 25.10-2 arm64 NVIDIA HPC SDK repository configuration files
ii libnvidia-cfg1-580:arm64 580.159.03-0ubuntu0.24.04.1 arm64 NVIDIA binary OpenGL/GLX configuration library
ii libnvidia-common-580 580.159.03-0ubuntu0.24.04.1 all Shared files used by the NVIDIA libraries
ii libnvidia-compute-580:arm64 580.159.03-0ubuntu0.24.04.1 arm64 NVIDIA libcompute package
ii libnvidia-container-tools 1.19.1-1 arm64 NVIDIA container runtime library (command-line tools)
ii libnvidia-container1:arm64 1.19.1-1 arm64 NVIDIA container runtime library
ii libnvidia-decode-580:arm64 580.159.03-0ubuntu0.24.04.1 arm64 NVIDIA Video Decoding runtime libraries
ii libnvidia-egl-wayland1:arm64 1:1.1.21-1ubuntu1 arm64 Wayland EGL External Platform library – shared library
ii libnvidia-encode-580:arm64 580.159.03-0ubuntu0.24.04.1 arm64 NVENC Video Encoding runtime library
ii libnvidia-extra-580:arm64 580.159.03-0ubuntu0.24.04.1 arm64 Extra libraries for the NVIDIA driver
ii libnvidia-fbc1-580:arm64 580.159.03-0ubuntu0.24.04.1 arm64 NVIDIA OpenGL-based Framebuffer Capture runtime library
ii libnvidia-gl-580:arm64 580.159.03-0ubuntu0.24.04.1 arm64 NVIDIA OpenGL/GLX/EGL/GLES GLVND libraries and Vulkan ICD
ii linux-headers-6.11.0-1014-nvidia 6.11.0-1014.14 arm64 Linux kernel headers for version 6.11.0 on ARMv8 SMP
It appears I’m already on the version of the gpu driver you recommended. If there’s anything newer that would help let me know. I’d much rather have the app crash than to need to physically unplug the box.
Thanks for the detailed report. The symptoms you’re describing—system hangs under moderate load with power delivery warnings from mlx5_core (27W insufficient power on PCIe slots)—suggest a potential interaction between power delivery, thermal management, and firmware behavior. Since you’ve confirmed firmware is up to date and already have a support ticket open, the next step is to gather diagnostic data to distinguish between power delivery, thermal, firmware, or hardware causes.
Please collect and share the following during or immediately after a hang event:
Thermal logs: CPU, GPU, and system temperatures at the time of hang (use sensors or BMC logs if available)
Power monitoring output: BMC logs or power delivery state during load (check ipmitool output if accessible)
Complete kernel logs: Run sudo journalctl -b > kernel_logs.txt after a hang to capture thermal throttling or power management messages
GPU state snapshot: Output of nvidia-smi -q and nvidia-smi dmon during load, showing GPU utilization and memory usage before the hang
Workload details: Which applications trigger the hang, and what is the load profile (GPU utilization %, memory usage %, duration before hang)
This data will help the support team determine whether the issue is related to power delivery limits, thermal throttling, firmware behavior under load, or a hardware defect.
Does it hang or powers down? If you have it connected to an Ethernet switch, you can check if the corresponding LED on the switch is on, otherwise you can check if there is an airflow in front or behind the unit.
I recently have the same issue when I do rapids task to my gb10, and seems like it has OOM as no response to ssh or any other services (Hermes agent down…etc), but I can ping the machine, so would like to know how to extract the report and see if I am having the same issue, btw, I have updated to latest.