Room temperature is 18 degrees, DGXSPARK running COMFYUI to generate video (about 10 minutes), often turns off automatically due to overheating. Now I dare not use COMFYUI, how to set the cooling fan.

This is what I do too. I tried many options but the most stable option is setting a cap at 2200. Under load, this is the only option survives.

Providing additional cooling (in my case placing the DGX Spark directly into the airflow of my air-con) gets rid of the issue, not shutting the device down even with full clock speeds while reducing the reported CPU temperature by roughly 5°C. This allowed me to compare performance between full and reduced clock speeds doing 7 image generation runs: Full Speed Reduced Speed [07:10<00:00, 21.52s/it] [07:23<00:00, 22.16s/it] [07:11<00:00, 21.55s/it] [07:23<00:00, 22.16s/it] [07:10<00:00, 21.53s/it] [07:25<00:00, 22.25s/it] [07:09<00:00, 21.50s/it] [07:24<00:00, 22.22s/it] [07:11<00:00, 21.56s/it] [07:29<00:00, 22.47s/it] [07:14<00:00, 21.75s/it] [07:36<00:00, 22.83s/it] [07:13<00:00, 21.68s/it] [07:30<00:00, 22.54s/it] Interestingly that means the GPU clock speed reduction to 2200 MHz only causes a roughly 3-4% performance loss and therefore is okay for me to keep using the device. Still, is there an official recommendation by the NVIDIA team on how we should handle this? If it is clear that replacing the device will fix the issue and there will be no software update to address this - I would love to RMA the device to avoid having to reduce the clock speeds.

No imagination required. Just gpu_burn developer@sparky : ~/github/gpu-burn $ ./gpu_burn 300 Using compare file: compare.fatbin Burning for 300 seconds. GPU 0: NVIDIA GB10 (UUID: GPU-cf0ef592-0423-ee7e-c6ee-65276769d5a0) cuInit returned 0 (no error) Initialized device 0 with 124610 MB of memory (11610 MB available, using 10449 MB of it), using FLOATS Results are 268435456 bytes each, thus performing 38 iterations 10.7% proc’d: 532 (18765 Gflop/s) errors: 0 temps: 83 C Summary at: Mon Jun 15 09:27:07 AM CEST 2026 21.0% proc’d: 1026 (16842 Gflop/s) errors: 0 temps: 78 C Summary at: Mon Jun 15 09:27:38 AM CEST 2026 31.3% proc’d: 1520 (17605 Gflop/s) errors: 0 temps: 85 C Summary at: Mon Jun 15 09:28:09 AM CEST 2026 38.7% proc’d: 1862 (17127 Gflop/s) errors: 0 temps: 85 C

My DGX spark immediately overheats when it is power on. I just have enough time to login then the screen goes grey pixelated then blank black screen and unit locks up. all I have is the mac address: 4c:bb:47:2f:34:92

My unit suffers a similar overheating issue. I am currently controlling the temp with (thanks to the advice above) sudo nvidia-smi -pm 1 # Enable Persistence Mode sudo nvidia-smi -lgc 300,2200 # Lock GPU Clocks DGX FIELD DIAGNOSTIC ./partnerdiag --field returns a Failed the PowerStress section. MODS-020000600139 | PowerStress | custommods | | Power | | Acceptable temperature limits exceeded or the thermal sensor is broken or miscalibrated This seems like an RMA qualify case. But my question is: whether an RMA is worth it or this issue can be resolved with future firmware update? As returning the unit risking 1) a few weeks without a working machine. 2) receiving a refurbished unit. 3) the replacement unit still have thermal problem persists as it is widely reported in the forum.

Check my tests here: [image] Cooler GB10 Temps - (almost) no performance lost DGX Spark / GB10 So a few times Iowered the clocks of my GB10 during long tasks to reduce the temps and I honestly never felt a lot (if any) performance lost, so today, I decided to actually measure: [image] I ran the same benchmark twice (codeneedle) in this case with Qwen3.6-35b-a3b-NVFP4. First run, normal clocks (hovered between 2400 and 2470mhz) and the 2nd run by limiting the clocks to 2000Mhz. Run 1 - Stock speed: GPU Power = 43w Speed (avg) = ~113tk/s you can see from the graph the orange line osc… I have my DGX permanently set to 2000Mhz now.

Both of my Sparks have to be downclocked to 2200 to be rock-stable in a cluster. However, when running solo, I can push the first one to 2400 and the second to 2250 before they fail DGX Field Diagnostics. The issue seems related to temperature. After a complete cooldown, both Sparks can pass the diagnostics, but they fail on the second or the third run. I am not a chip engineer, but this behavior suggests that the stock clock settings might be too high, as there is a clear variance between chips in how far they can be pushed. On the Raspberry Pi forums, this is often called the “silicon lottery”—lucky ones can overclock their Pi by up to 20%, while others can’t overclock past the default settings at all. My private conspiracy theory is that the DGX Sparks were originally designed for 2000–2200 clocks, but at the last minute, a decision was made to overclock them to 2450 by default. Otherwise, I cannot imagine how this thermal instability wasn’t caught during development. Literally, all it takes is running any model in a cluster, setting the GPU memory to 0.9, executing llama-benchy with a concurrency of 8, and waiting 15–20 minutes until one of the Sparks hangs. For now, I’ve decided not to start an RMA because I don’t see any other issues, and at 2200, both Sparks work reliably 24/7.

I did some tests and decided to leave my spark like this: built a ComfyUI workflow with WAN22 video gen + ONNX model per-frame upscale . WAN22 diffusion keeps the GPU under sustained load, while ONNX upscale slams it with short burst responses. loop the workflow with various clocks 2000-2350 (default 2418 just can’t pass). so far capping the gpu clock at 2260 produces optimal workflow time. keeps the gpu sits barely over 80c (no outside cooling). I leave the workflow loop overnight. workflow looped all night, no thermal shutdown; only ~3% drift in loop time. Yes, the FE Spark left performance room for OEM partners. I guess the FE enclosure is not originally designed to house a GB10 chip that needs to be OC to get 1 petaflop :) sudo nvidia-smi -lgc 300,2260

@NVES the DGX Spark dashboard already includes a system memory and a gpu utilisation monitor. would be possible to add a temperature monitoring?[feature-request]

I am experiencing a very similar issue. In my department, we have 8 units of GB10, and 2 of them are exhibiting this exact problem where they overheat out of nowhere while completely idle. I believe all of our GB10 units have been updated to the latest version via the Dashboard, yet the issue persists on these two. Even with no software running on those affected units, the chassis gets incredibly hot—to the point where I can’t keep my hand on it for more than 3 seconds. However, running nvidia-smi only reports a temperature of 51–54°C, which clearly doesn’t match how hot the hardware physically is. Upon closer inspection, I found that the onboard fan was completely stationary. Most of the time the fan works normally, but it occasionally fails and stops spinning under certain unknown conditions. I managed to temporarily solve this by placing an external 12cm DC fan to blow at a very low speed toward the unit, which successfully brought the temperature down. It seems there are two main issues here: Inaccurate temperature readings / telemetry mismatch Intermittent fan stalling at idle I am very concerned that the other 6 remaining GB10 units might suddenly experience this overheating issue as well. Has anyone found a confirmed solution or workaround for this? Any insights would be greatly appreciated.

DGXSPARK temperature too high, automatic shutdown。

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

azampatti June 17, 2026, 2:10pm 171

Check my tests here:

I have my DGX permanently set to 2000Mhz now.

Topic		Replies	Views
Spark abruptly shuts down DGX Spark / GB10 dgx-spark-issue	47	4300	October 9, 2026
DGX Spark (GB10) reproducibly hard powers-off under GPU load — fully updated, zero crash capture DGX Spark / GB10 boot , kernel , ota , dgx-spark-issue	13	1624	June 14, 2026
DGX Spark. low fan speed, high temps, device very hot DGX Spark / GB10 kernel , gpu , fan-facts , debugging-and-troubleshooting	62	7519	October 27, 2026
Dgx spark shut down without rebooting DGX Spark / GB10	21	1192	May 20, 2026
DGX Spark stability / out of RAM / overheating DGX Spark / GB10 llama , dgx-spark-issue	51	5053	September 7, 2026
DGX Spark failing ThermalStress Fieldiag DGX Spark / GB10 dgx-spark-issue	15	446	October 15, 2026
DGX Spark GB10: partnerdiag PowerStress reproducibly hard-powers-off the box — acpitz 88→97.8°C in 5s, all other field tests pass DGX Spark / GB10 Projects dgx-spark-issue	9	948	July 24, 2026
DGX Spark reaches ~95°C thermal-zone readings and has abrupt outages under 122B inference — could cooling or USB-C PD be involved? DGX Spark / GB10	12	674	October 20, 2026
DGX Spark fans stop when the screen goes dark or running from SSH — box gets too hot to touch* FIRE HAZARD DGX Spark / GB10 dgx-spark-issue	19	2237	October 28, 2026
DGX Spark constantly shut down after first set up DGX Spark / GB10	8	791	January 11, 2026

DGXSPARK temperature too high, automatic shutdown。

Related topics