Honestly surprised: Intel GPU draws only ~3W for the exact same vision workload we currently run on an Nvidia RTX 5060 Ti at ~70W
Tested this K13 mini-PC machine and were honestly surprised - for the same vision workload we currently run on an NVIDIA RTX 5060 Ti at around 70W, the Intel GPU was drawing only about 3W.
The workload:
- 20 H.264 camera streams
- hardware video decode
- YOLO object detection
- everything running on the GPU
On NVIDIA, we use NVDEC for decode and CUDA for YOLO.
On this Intel, the pipeline runs through Quick Sync Video / VA-API for decode and Intel GPU inference for YOLO.
And the little machine has solid cooling too: around 37C at idle and stable around 70C under the above load.
To be clear, this is not LLM inference. This is the “pre-LLM” stage for physical-world AI systems: processing video streams, detecting objects, and turning camera feeds into structured signals before an LLM or agent reasons over them.
For this kind of workload, Intel’s performance per watt looks surprisingly good. We’re still testing, but this changed how I think about small, low-power edge AI boxes for camera-heavy workloads.
The box has an Intel Core Ultra 7 256V with Intel Arc 140V graphics. I’m running Reefy.ai OS on it, which reports these GPU/video/thermal metrics out of the box.
Update:
I ran the wall-power experiment proposed in the comments - appreciate the push!
For the same task:
Mini PC: 17.5 W delta
Nvidia GPU machine: 130.4 W delta
So yes - the difference is still big, but not as dramatic as my original estimate. Measured ratio is about 7.5:1, not 23.3:1.