Status and Experience on Thermal Performance

Here’s my take on cooling of Gigabyte AI Top Atom:

Version 1: Basic laptop cooler stand , with a table top fan in the front. Almost regular 24hrs running qwen2.5:32b-instruct , temperatures revolve around 90-92 Degree Celsius.

Running ComfyUI WAN2.2 Remix 12V Workflow, around 97-99 Degree Celsius.
System would often crash at heavy ComfyUI workloads.

Version 2: DGX COOLER ULTRA MAX PRO ++++ (Overkill)

Front : 3 x 40mm FAN

Bottom : 4 x 120mm FAN

Rear: 2 x 80mm FAN

With regular 24hrs running qwen2.5:32b-instruct , temperatures revolve around 85-88 Degree Celsius. Also, I saw no crashes at all with continues ComfyUI workloads.

Running ComfyUI WAN2.2 Remix 12V Workflow, around 92-96 Degree Celsius. System would not crash, I ran 10 I2V Workflows in queue.

I turned off my DGX COOLER ULTRA MAX PRO ++++ mid ComfyUI WAN2.2 Workflow, and the system crashed in just 1 min.

The setup seemed to be working perfectly fine, but I think it’s overkill.

Version 3:

Front: 140mm FAN

Bottom: 120mm FAN
Top: 120mm FAN

Rear: 140mm FAN (With an exhaust cover)

Temperatures revolve around 75-80 Degree Celsius running qwen2.5:32b-instruct.

Running ComfyUI WAN2.2 Remix 12V Workflow, max temperature it hits is 97 Degree Celsius. No crashes at all.

The fans are running at max RPM, no noise at all, all fans are Noctua Industrial.

I 3D built & printed Version 2 & 3 in house with a PLA-CF Filament, would have preferred PETG-CF , but did not have enough of it in stock with me.

VERDICT: Version 2 seemed better at first, it gave server level cooling, but the smaller 40mm Fans made way too much noise, almost server level noise. I added a PWM Fan controller, so found a sweeter spot, where the 40mm fans were making slight less sound. Maybe Noctua 40mm fans would make less noise, not sure. Have ordered Noctua NF-A4x20 PWM , let’s see if they make less noise.

For my purpose and the rack space I have, Version 3 is much more suited to me, it fits perfect in my custom rack shelves space.

Also, what I learned, its best to direct the airflow for the system to work at the best temperatures. Directing the rear hot air with proper exhaust duct airflow seemed to be cooling the system much better.

I have added an option to add another fan mount option on the top as well on Version 3 , so it can cool the box from the top as well.

I am located in India. Outside temperature currently is around 34-36 Degree Celsius. I made all the tests without any Air Conditioning in the room.

Would be happy to share the 3D stl files if anyone needs them.
Printed all these parts in my Bambu Lab P2S.

I think full load depends on how bigger of a LLM model you have loaded, your GPU can still be utilizing 98% of VRAM , but with a smaller model loaded, and in that case the temps won’t reach the peaks. Load a model with around 50-70 GB , and then run the model with a high KV Cashe prompt, and see the temps. Maybe with heavy pre compute prompt.