[DGX Spark] Overheating, 100W Power Cap, 30W Safety Mode — Complete Diagnostic Guide

Plain-Language Version: When Your AI Machine Is Secretly Running on Fumes

The NVIDIA DGX Spark (and its ASUS GX10 variant) is a desktop AI workstation that needs a lot of power — up to 180 watts delivered through a USB-C cable. Think of it like a high-performance car: it needs premium fuel at full pressure to perform.

Some units shipped with a defective power negotiation chip. The machine turns on, looks normal, and even runs AI tasks — just painfully slowly. It is like driving a sports car with the fuel line pinched: the engine runs, but it never gets past first gear. The system caps itself at about 30 watts instead of the full 180.

The tricky part is that everything looks like a software problem. You spend hours tweaking AI model settings, swapping configurations, and reading logs — when the real issue is that the hardware is starving for electricity. One terminal command (nvidia-smi) reveals whether you are affected in 30 seconds.

If you are, the fix is not software — it is a warranty replacement (RMA). This article shows exactly how to diagnose it and what to do next.

---

The Diagnostic (Run This First)

Fire any GPU-bound workload — an inference request, a quick matrix multiply, anything — then sample:

nvidia-smi --query-gpu=power.draw,utilization.gpu,clocks.sm --format=csv,noheader

Two outcomes:

Healthy:

35.65 W, 96 %, 2522 MHz

30W safety mode:

4.80 W, 2 %, 2411 MHz

In safety mode, GPU utilization stays near zero regardless of how much work you throw at it. The machine is not thermal-throttling — it is power-starved at the PMIC level.

On my unit, running Qwen3.5-35B-A3B-FP8 via vLLM during a 300-token generation:

power.draw = 35.65 W  |  utilization.gpu = 96%  |  clocks.sm = 2522 MHz
throughput  = ~50 tok/s

No throttling. Normal behavior.

Root Cause: PD Controller Firmware Negotiation Failure

The GX10 takes power via USB-C PD 3.1 (180W EPR spec). The PD controller negotiates with the 240W adapter to unlock high-power mode — somewhere around 20V/5A or 28V.

On affected units, this negotiation fails silently. The adapter is plugged in, the machine boots normally, but the PD controller never escalates beyond the default safe power level. The PMIC then restricts the entire main power rail to approximately 30W.

The dmesg signature:

Detected insufficient power on the PCIe slot (27W)

That message appears across multiple Mellanox network controllers — confirming this is system-wide power starvation, not a single component issue.

What Does Not Fix It

ASUS's recommended procedure is a "double flash" of the PD controller firmware:

Flash → Reboot → Flash → Reboot

Users in the NVIDIA Developer Forums thread followed this exactly. The throttling persisted. NVIDIA support's conclusion: hardware defect, RMA required.

This is not a fixable firmware issue. The PD controller itself has failed in a way that software cannot reach.

The Other Power Bug (Different Problem, Actually Fixable)

There is a separate issue worth knowing: GPU stuck at 5W with 0% utilization, caused by running an outdated driver stack (550.54.15 + CUDA 12.4).

This one is software:

sudo apt dist-upgrade
sudo fwupdmgr refresh && sudo fwupdmgr update

Upgrading to Driver 580.x + CUDA 13.0 resolved it. Confirmed fixed as of January 2026.

If your GPU utilization is low but not stuck at near-zero, check driver versions before assuming hardware failure.

Firmware Version

The latest firmware as of this writing: BIOS v0103 (2026/03/18)

SOC / 0x305
EC  / 2.78.18.3
PD  / 0x507

My unit is currently on PD 0x500. Updating is worth doing, but if you are not seeing 30W symptoms, it is not urgent. If you are in safety mode, updating PD firmware is unlikely to help — but do it before initiating RMA to rule it out.

2026 Community Controversy: Carmack + the 100W Problem

In April 2026, John Carmack publicly criticized the DGX Spark for falling far short of the advertised 1 PFLOPS sparse FP4 performance. The NVIDIA developer forums were subsequently flooded with reports of three categories of failure:

Symptom	Cause	Severity	Fixable?
Power capped at ~30W, 0% GPU utilization	PD controller hardware defect	Critical	No — RMA required
Power capped at ~100W (less than half of 240W rating)	Insufficient cooling → thermal throttling	High	Partially — improve airflow
GPU shows 5W / 0%, but dmesg is clean	Driver 550.x bug	Low	Yes — upgrade driver

The 100W issue is distinct from the 30W issue. 30W is a PD controller manufacturing defect (permanent). 100W is thermal throttling (environmental). If your machine's power draw plateaus around 100W under load and GPU temperature is near the limit:

nvidia-smi --query-gpu=power.draw,temperature.gpu,clocks.throttle_reasons.sw_thermal_slowdown --format=csv,noheader

Active in the thermal slowdown column confirms throttling. Mitigations:

• Ensure intake vents are unobstructed
• Keep ambient temperature below 25C
• Consider an external fan aimed at the chassis underside
• Vertical orientation improves convection vs. horizontal

Note: NVIDIA's CES 2026 software update claims 2.5x performance improvements. If you are on an older DGX OS version, update before judging.

What Was Gained

What cost the most time: The 30W issue manifests silently. The machine boots, runs commands, serves requests — just slowly. Without the direct power.draw + utilization.gpu check, it would look like a vLLM config problem or an underpowered model.

Transferable diagnostic: For any unexpectedly slow inference on new hardware, sample power.draw and utilization.gpu simultaneously under load. If utilization is high and power is low relative to rated TDP, the power delivery chain deserves scrutiny before any software tuning.

The pattern: Hardware defects that look like configuration problems waste the most time because the debugging surface is infinite on the software side.

Checklist

1. Run nvidia-smi --query-gpu=power.draw,utilization.gpu,clocks.sm --format=csv,noheader under load
2. If utilization is low: check driver version, upgrade to 580.x + CUDA 13.0
3. If utilization is high but throughput is inexplicably low: check for 30W dmesg signature
4. If 30W confirmed: double-flash PD firmware per ASUS docs, then RMA if no change
5. Update to BIOS v0103 (PD 0x507) regardless — cooler temps reported by other users post-update

---

Also in this series:

• Part 5 — FP8 KV Cache Repetition
• Part 4 — NVFP4 Debugging on SM121

Related deep dive:

• LLM Deep Dive Part 2 — TurboQuant KV Cache Benchmark