I run a DGX Spark as an always-on box for local LLM inference. A couple of weeks ago it
started being off in the mornings — not asleep or hung, powered down, with no reboot and
nothing on screen. It’s happened four times in five days, always inside the window where my
scheduled GPU workload runs, never while idle.
There’s no forensic trace at all: journal stops mid-line, no panic or oops, kdump produces
no vmcore, /sys/fs/pstore is empty, no Xid, no thermal trip. That reads as power loss, not
a crash. So I ran NVIDIA’s field diagnostic to see if I could reproduce it on demand — and
I can, every time, on PowerStress.
Details:::
System
Product NVIDIA_DGX_Spark, Product Version A.7
SBIOS 5.36_0ACUM018
Driver 580.159.03 (open kernel module, aarch64)
Kernel 6.17.0-1021-nvidia
partnerdiag r9.257.8
Everything except PowerStress passes clean
From summary.json — four tests, all Notes: “OK”, error code …000000000000:
┌────────────┬──────────┬────────┐
│ Test │ Duration │ Result │
├────────────┼──────────┼────────┤
│ GpuStress │ 200.1s │ PASS │
├────────────┼──────────┼────────┤
│ C2CStress │ 6.6s │ PASS │
├────────────┼──────────┼────────┤
│ CpuStress1 │ 8.7s │ PASS │
├────────────┼──────────┼────────┤
│ CpuStress2 │ 602.7s │ PASS │
└────────────┴──────────┴────────┘
PowerStress never returns a result — the machine powers off mid-test
This is the finding. PowerStress produces no error code because the box dies before it can
write one. From test_status.log:
{
“name”: “PowerStress”,
“startTime”: “2026-07-18 11:01:41.248446-0700”,
“timeoutInSecs”: 1800,
“testSequenceNum”: 4,
“statusIndex”: 4
}
No endTime, no durationInSecs, no entry in summary.json. Started 11:01:41, machine dead
~11:05 — about 3m20s in.
Reproduced with external instrumentation
Since partnerdiag’s own logging dies with the box, I re-ran PowerStress alone with a 1 Hz
thermal sampler running outside it, writing to disk continuously. Started 13:37:26, dead
13:43:38 — 6m12s. The sampler caught the event (all zones acpitz, °C):
13:43:27 zone0=88.7 zone2=67.5 zone4=85.1 ← stable plateau ~6 min
13:43:30 zone0=88.8 zone2=67.4 zone4=85.6
13:43:31 zone0=88.0 zone2=73.2 zone4=76.5 ← discontinuity
13:43:32 zone0=93.9 zone2=79.5 zone4=69.7
13:43:33 zone0=96.8 zone2=81.9 zone4=69.6
13:43:36 zone0=97.8 zone2=83.6 zone4=70.1 ← +9.8°C in 4s
13:43:38 zone0=96.5 zone2=83.4 zone4=70.2 ← last sample, power gone
zone0/zone5 hold a flat 88.7°C plateau for six minutes, then jump nearly 10°C in four
seconds while zone4 simultaneously collapses from 85.6 to ~70 and stays there. A 9°C rise
in 2s isn’t thermal mass. That looks like either a sensor handoff/miscalibration or a real
unmanaged hotspot that nothing is reacting to.
Ruled out
- Not software — reproducible on demand with NVIDIA’s own diagnostic, with my entire stack
stopped and the system in multi-user.target. - Not a kernel crash — no vmcore, empty pstore, and a log file left NUL-padded (filesystem
extended it, data never flushed), which only happens on abrupt power loss. - Not thermal throttling working as intended — nothing trips, the box just goes.
- journalctl --list-boots under-reports the death time (journald is down during the
isolate); trust the external sampler.
Same signature as several existing reports
- DGX Spark (GB10) reproducibly hard powers-off under GPU load — fully updated, zero crash capture — identical SBIOS, driver, kernel, empty
pstore - DGX Spark hangs under vLLM load, fieldiag fails on the thermal sensor
- Total host freeze (not process hang) during multi-node TP=2 vLLM prefill on 2× DGX Spark GB10, zero forensic trace across kdump/watchdogs/netconsole
Ask
This looks like error class 082-000-1-020000021139 (temperature limits exceeded or thermal
sensor broken/miscalibrated), but I can’t produce that code because the test never
completes. Is that sufficient to open an RMA? Happy to provide the full log bundle, the
sampler CSV, or run any specific test — with the caveat that every PowerStress run is a
hard power-off on a machine I depend on, so I’d rather not repeat it without a reason.
Serial available on request.