Deskpi t0, tt, t1,t2 with fans? A datacenter rack is optimized for removing heat. Deskpi makes similar design for homelabs. Theres already best practices for the most effective heats removal.
Didn’t read the thread but though it might be usefull
Just measured overall power consumption (from wall socket 120V) on my GX10 under LLM load (llama.cpp doing prompt processing with qwen3.6-27b):
normal clocks (after “nvidia-smi -rgc”): nvtop shows up 85-95W, total power consumption is up to 180W
2100MHz clock (“nvidia-smi --lock-gpu-clocks=200,2100”): nvtop - 55W, total - 125W
Idle state, nvtop - 4W, total - 26W
“15W power locked” bug state: nvtop - 12W, total - 62W
This is even without serious network load in headless mode (only ssh session over ethernet), no anything else connected.
Are you all actually getting high temps out of these boxes? I have an ASUS variant, and for extremely long sessions of document organization it seems to be at about 61-65C. Barely anything. Maybe it spikes when doing compilation or other mixed usage tasks.
I guess a fan wouldn’t hurt either way, though. Most of my noctua fans are busy in other machines.
Mine is 35c at idle, but goes to around 90c under load!
Tue Jun 16 09:06:47 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.159.03 Driver Version: 580.159.03 CUDA Version: 13.0 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GB10 On | 0000000F:01:00.0 Off | N/A |
| N/A 61C P0 34W / N/A | Not Supported | 95% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 3880 G /usr/lib/xorg/Xorg 56MiB |
| 0 N/A N/A 4046 G /usr/bin/gnome-shell 52MiB |
| 0 N/A N/A 157419 C VLLM::EngineCore 11814... |
+-----------------------------------------------------------------------------------------+
Am starting to wonder if I am doing something wrong here. It’s barely using any wattage, at least according to Nvidia-SMI, and the temp only hovers around that range with 122B chugging away constantly in vllm.
Edit: Testing looks fine, though…
sparkler@spark-b60e:~/spark-gpu-throttle-check$ python3 spark-gpu-throttle-check.py
============================================================
Spark GPU Throttle Check
============================================================
GPU state at idle:
Clock: 2483 / 3003 MHz
P-state: P0
Power: 12.8 W
Warming up GPU (2.0s)...
Collecting 20 samples under load (0.5s interval)...
Threshold: 1400 MHz
# Clock (MHz) Max (MHz) PState Power (W)
───── ─────────── ───────── ────── ─────────
1 2457 3003 P0 82.2
2 2457 3003 P0 82.2
3 2457 3003 P0 82.0
4 2457 3003 P0 82.2
5 2457 3003 P0 82.4
6 2457 3003 P0 82.3
7 2457 3003 P0 82.5
8 2457 3003 P0 83.2
9 2457 3003 P0 82.8
10 2457 3003 P0 81.8
11 2457 3003 P0 82.7
12 2457 3003 P0 83.6
13 2457 3003 P0 83.0
14 2457 3003 P0 82.3
15 2457 3003 P0 83.1
16 2457 3003 P0 83.8
17 2457 3003 P0 83.2
18 2457 3003 P0 82.6
19 2457 3003 P0 83.3
20 2457 3003 P0 83.8
────────────────────────────────────────────────────────────
RESULTS
────────────────────────────────────────────────────────────
Samples: 20
Peak clock: 2457 MHz
Average clock: 2457 MHz
Avg power draw: 82.8 W
Below threshold: 0% of samples < 1400 MHz
┌────────────────────────────────────────────────────────┐
│ PASS — GPU clocks look healthy under load. │
│ Peak: 2457 MHz, Avg: 2457 MHz │
└────────────────────────────────────────────────────────┘
Are you just processing one request at a time? Mine is mostly 35W if it’s only processing one request too, but it goes up if I run a benchmark with 8-10 concurrent requests. In both cases, it shows around 100% GPU utilisation.
Yeah, if I actually tried doing multiple of my current requests at the same time on this model, I think it would just run out of context very quickly. That makes sense, so I guess that’s what it was. I bought one of the 140mm fans anyway just in case. IME larger size generally means quieter for the same amount of flow, though it depends on the actual of the spark’s vent whether this is efficient or not.
That said, it still only spiked to about 60C even while running that stress test, so that’s kind of weird. The stress test did pin it at 80+W.
As far as cooling the adapter, not sure how effective that would be tbh. It’s made out of plastic, right? I doubt it’s actually radiating that much of its internal heat outwards. Almost all of it is internal. The heat transfer coefficient of plastic is probably extremely low?
Just want to report that my choice to get the 140mm fan instead (same Amazon page) was a pretty good one. Same price and it almost perfectly fits the ASUS variant and essentially serves as a nice stand for it. Pretty supportive, I don’t see the need to print out any bracket. As far as my temps for just what I’m doing, went down from 61-63C to about 55-56, so not bad. It’s not too loud at all, either. Difficult to hear over ambient.
I wouldn’t really worry about trying to just cram down all of the CFM that this thing is outputting up that little hatch. I think most of the benefit is going to come from giving the ASUS fan more ground clearance and access to more cooler air. I don’t think the airflow works in such a way that if you just push more directly in there, it’ll mean more cooling. In fact, I would kind of be worried about trying to cram more air past a fan to begin, with the way these fans work.
You are correct if you have a fan blowing from the front of the Asus GX10.
However, with the Spice Harvester cage, it simply creates air pressure which, without any easier path (hence the differences between sealed vs just blowing), will go through the 150x100 intake. The fan size matters less than the CFM, a bigger fan at the same RPM will generate more air flow compared to a smaller fan, but you can compensate with a higher RPM (and higher “jet engine” noise). 120x120 seems sufficient. A blower or turbo fan might be better than the axle fan that is currently designed for, but also harder to source and get the correct dimensions.
The same fan at 2000 rpm will create 4 times the pressure of a fan at 1000 rpm, so ideally a temperature control fan to maintain the temp below throttle threshold rather than running the fan at max speed in my opinion would be better .
I tends to prefer a intake rather than a exhaust fan but I send out for 3D printed the exhaust design before the new intake design was available.
A intake fan will push air in via a single intake, if you increase the RPM, the air might leaks , which should be fine as it would push the air out via any gaps.
But if you have a exhaust fan, and you increase the RPM, this will create a low pressure zone, and the air will come in via “gaps”, and these could potentially include unused (and used) HDMI, USB, etc. ports. Dust eventually will get stuck at this gaps.
I did a brief experiment using “Flydigi BS1 Laptop cooler”, which is a intake fan. It comes with a BLE fan controller, so I create a python script that runs it at 1000 rpm during idle and spins up to 3000 rpm. During my load testing and benchmarking, it never actually need to reach 3000 rpm, maximum was probably 2300 to 2500 which maintain the temp at 10 degrees headroom.
It works out pretty well and prevented thermal throttling in all my benchmarks and load testing, maintaining at 10 degrees below throttling, except it looks horrible to have a “mini PC” GX10 sitting on top on a laptop cooler, AND it’s too expensive a gadget to gut just for the controller although I’m quite tempted.
So, if anybody knows how I can get hold of a BLE / BT 12V fan controller cheaply, let me know. But for now, Dr Thomas design is probably the best option.
Actually it does. Quite a few guys has done that testing here. You don’t need to lower temp by xx degrees, you just need to keep it below throttling with sufficient headroom.
You get THAT benefit by just lying the ASUS on it’s side.
I am both proud and ashamed to report this – after mulling for months on how best to cool my double GX10 with a 3D-printed design, I just went full stupid. And it worked.
2 Arctic P14 Pro PST fans, a 12V PVM fan controller, 4 zipties, a cardboard box, and an old mousepad for a base.
My GX10s are 20-25 degrees cooler now.
