I have a Asus Ascent GX10. Installed the latest BIOS and firmware available as of June 27, 2026. I am running benchmarks that stress the GPU to it’s max. I am only drawing 60 watts and TFLOPS is about 65. The CPU is not busy during the benchmark so I would expect close to the full 140W for the GPU. I have done the “disconnect the power supply from the wall and wait 60 seconds” method a number of times. I am using Claude Code to assist and it is convinced it is a firmware/platform cap somewhere.
Here is what Claude has to say about it :
Two things stand out here:
SW Power Cap: Not Active — a software power cap mechanism exists, currently off
SW Power Capping: 2858182 us — the GPU has accumulated ~2.9 seconds of SW power cap throttle time (almost certainly from our benchmark runs)
That second one is significant — let me dig into it and check what throttle reasons are available:
-pl (power limit) — the classic nvidia-smi -pl command exists, but all power limits show N/A meaning the driver doesn’t expose a settable limit for this chip
power-profiles — the newer workload power profiles API (pre-tuned recipes) is explicitly “not supported on this device”
SW Power Capping counter — the GPU has been throttled by a software power cap during our runs, but it’s implemented deep in the driver/firmware with no user-facing control knob
The SW Power Capping: 2858182 us is the smoking gun — the cap is real and it’s being enforced in software, but NVIDIA has not exposed any API to adjust it on the GB10. This is almost certainly a platform-level TDP policy baked into the driver for the DGX Spark form factor, which is exactly why the support ticket is the right next move — only NVIDIA can change this.
Have you tried unplugging it completely for a few min and then trying again? Mine slowed after exceeding, memory crashing, and rebooting. Unplugging restored the original speed.
Isn’t 140W max limit for whole chip (both CPU die + GPU die)? I’ve seen 85-90W max on GPU power at some task but for quite short time, pretty sure this is the limit for the GPU.
According to Nvidia the total CPU+GPU power is 140W, but that does not mean GPU can utilize the full 140W envelope. From personal experience, normal model runs over vLLM push GPU power to 35-45W range. Some extremely optimized tensor operations and burn tests can hit 90+W, which I believe is the maximum we can get out of the GPU. I only tested it while passing so unfortunately cannot provide the exact circumstances that drew 90+W.
It would be helpful to know exactly which benchmark tools or projects were used for the testing. If possible, I’d also appreciate seeing the benchmark logs.
Also, it’s worth noting that many benchmarks—especially those involving LLMs—often do not fully utilize the GPU’s power. In some cases, I’ve seen the GPU running at over 2500 MHz while drawing less than 70W during benchmark runs.
If possible, I recommend testing the Wan 2.2 i2v workflow included in the ComfyUI templates. This workflow is extremely demanding and represents a workload that even NVIDIA’s official Field Diagnostics and other intensive benchmarking tools often fail to reproduce.
If the cooling system is working properly, it’s possible to observe power consumption exceeding 100W during this workload.
Because of how demanding it is, many users have actually used this workflow to reproduce or detect system shutdown issues under heavy GPU load
Depending on the worload, 100% GPU utilization might render 30-90w of consumption. It all depends.
In my case, I use my GB10 down-clocked for less power consumption (lower temps) Cooler GB10 Temps - (almost) no performance lost , but with stock clocks, I see 90W on Prompt Processing and 45W in Token Generation.
But also depending on the model, this might change.
One important clarification first: on DGX Spark / GB10 systems, the published 140W figure refers to the GB10 SoC TDP, not a guaranteed GPU-only power target. So seeing less than 140W on the GPU by itself is not necessarily unexpected.
That said, the behavior you’re describing may overlap with a low-power or degraded-performance state that has been reported by some users, especially if performance is materially below normal and a full AC power removal does not restore expected behavior.
To help narrow down whether this is expected workload behavior or a degraded-state issue, could you please share the following:
The exact benchmark or workload you are running, including command line and options
A short power capture while the benchmark is running, for example:
nvidia-smi -q -d POWER -l 1
Confirmation that you are using the included 240W power adapter
If available, a Field Diagnostics result or log bundle
If you can attach those details, we can check whether this matches the known degraded-state behavior or if this looks like a different issue on the ASUS GX10 specifically.