DGX Spark hard-freezes under sustained few minutes inference; PowerStress thermal failure; support portal unavailable

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

My NVIDIA DGX Spark reproducibly hard-freezes under sustained GPU inference.

  • DGX OS 7.5.0
  • Kernel 6.17.0-1029-nvidia
  • Driver 580.173.02
  • EC 3.5.8; current firmware; fwupdmgr reports no updates
  • Original 240 W adapter verified
  • Display manager and internal airflow verified
  • GpuStress passes
  • PowerStress fails with MODS-020000610139: acceptable temperature limits exceeded or thermal sensor broken/miscalibrated
  • MiniMax-H3 864×480/5-second inference hard-freezes the complete system, without OOM, NVIDIA Xid, kernel panic, or application exception
  • ACPI temperature reached approximately 91–93°C
  • Idle temperatures were normal: GPU 47°C and hottest ACPI zone 49.8°C at approximately 4 W. During MiniMax-H3 inference at 96% GPU utilization and approximately 70–83 W, the GPU reached 84°C and the hottest ACPI/SoC zone reached 93.1°C within roughly two minutes. The system subsequently hard-froze.

Let me know what other info is required.

2

I’m not sure if my experience will be helpful, but I had similar symptoms and resolved the issue by downgrading the firmware to an earlier version and locking the maximum clock speed to 2000 MHz.

However, since this doesn’t solve the root cause, I think I’ll have to send it in for an RMA if the problem occurs again.

3

i have been thinking about downgrading, but it is supposed to be running long-sustained sessions like this without crashing, and managing the clock speed based on its load and temp. that should be fixed.

4

hey,

I have basically same specs (kernel,driver…) on asus , the 91+ temps just anecdotally seems high , Im testing MiniMax-H3 so I;ll drop a quick smoke test and monitor acpi…

$ echo "=== DGX Spark (GB10) — H3 T2V render thermal report ==="
echo ""
echo "Model: MiniMax H3 T2V (NVFP4), 864×480, 16 steps, 3s clip"
echo "Render time: 101s on single DGX Spark"
echo ""
echo "Temperatures:"
echo "                    Idle        Under load (101s)  Delta"
echo "GPU (die)          45°C         58°C              +13°C"
echo "ACPI (SoC)         47°C         62°C              +15°C"
echo "ConnectX-7 NICs    48°C         62°C              +14°C"
echo ""
echo "GPU power draw: 15.17W (under load)"
echo "GPU memory: ~24 GB (diffusion) + 15 GB (TE) + 5 GB (VAE) ≈ 44GB"
echo ""
echo "Output: 721 KB mp4, 3.0s, H.264 + AAC stereo audio (native)"
echo ""
echo "ACPI sensors report 7 thermal zones (Tdie, Tccd0, Tccd1, etc.)"
echo "All well below throttling thresholds; fan curve not aggressive enough"
echo "to recover to idle temps within 30s post-render — sustained workloads"
echo "would benefit from the PD wedge fix (see cheatsheet)."
=== DGX Spark (GB10) — H3 T2V render thermal report ===

Model: MiniMax H3 T2V (NVFP4), 864×480, 16 steps, 3s clip
Render time: 101s on single DGX Spark

Temperatures:
                    Idle        Under load (101s)  Delta
GPU (die)          45°C         58°C              +13°C
ACPI (SoC)         47°C         62°C              +15°C
ConnectX-7 NICs    48°C         62°C              +14°C

GPU power draw: 15.17W (under load)
GPU memory: ~24 GB (diffusion) + 15 GB (TE) + 5 GB (VAE) ≈ 44GB

Output: 721 KB mp4, 3.0s, H.264 + AAC stereo audio (native)

ACPI sensors report 7 thermal zones (Tdie, Tccd0, Tccd1, etc.)
All well below throttling thresholds; fan curve not aggressive enough
to recover to idle temps within 30s post-render — sustained workloads
would benefit from the PD wedge fix (see cheatsheet).
```

Just some info , hope it could be useful. Good luck.

6

Hi @tannerhaggerman,

Thank you for reporting this — a few questions and next steps:

Environment:

  • Where is your DGX Spark physically placed? Is it in an enclosed space (cabinet, shelf) or open area?
  • What is the approximate ambient temperature in the room?
  • Have you noticed the fan behavior before the freeze — does it spin up to max speed, or does it seem relatively quiet?

Field Diagnostics:
We’d like you to run the official NVIDIA Field Diagnostics to rule out a hardware issue. Here’s how:

sudo apt-get update
sudo apt-get install dgx-spark-fieldiag
cd /opt/nvidia/dgx-spark-fieldiag
sudo ./partnerdiag --field

The diagnostic takes approximately 30 minutes and will produce a clear PASS/FAIL result. If it reports a FAIL, that qualifies your unit for an RMA. Post the result here and we’ll authorize RMA via this thread, which you can then reference in an NVIDIA consumer support case.

The full guide is here: DGX Spark Field Diagnostics User Guide

Additionally, to help us gather more data before the next freeze:

  • Run sudo cat /sys/class/hwmon/hwmon*/temp*_input during your inference workload to capture all thermal zone readings just before the freeze
  • Check journalctl -b -1 | tail -50 after a reboot following a freeze to see if there’s any last-gasp kernel message

Thanks,
Neill

7

Hello again @tannerhaggerman,

Quick note: Something caught my eye after posting that I missed at first pass, if “PowerStress fails with MODS-020000610139: acceptable temperature limits exceeded or thermal sensor broken/miscalibrated” was the result of a field diagnostic output, can you please share the full output here for RMA approval?

Also, can you elaborate on “support portal unavailable”?

Thanks,
Neill